extern "C" magma_int_t
magma_didr(
magma_d_matrix A, magma_d_matrix b, magma_d_matrix *x,
magma_d_solver_par *solver_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_IDR;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
solver_par->init_res = 0.0;
solver_par->final_res = 0.0;
solver_par->iter_res = 0.0;
solver_par->runtime = 0.0;
// constants
const double c_zero = MAGMA_D_ZERO;
const double c_one = MAGMA_D_ONE;
const double c_n_one = MAGMA_D_NEG_ONE;
// internal user parameters
const magma_int_t smoothing = 1; // 0 = disable, 1 = enable
const double angle = 0.7; // [0-1]
// local variables
magma_int_t iseed[4] = {0, 0, 0, 1};
magma_int_t dof;
magma_int_t s;
magma_int_t distr;
magma_int_t k, i, sk;
magma_int_t innerflag;
double residual;
double nrm;
double nrmb;
double nrmr;
double nrmt;
double rho;
double om;
double tt;
double tr;
double gamma;
double alpha;
double mkk;
double fk;
// matrices and vectors
magma_d_matrix dxs = {Magma_CSR};
magma_d_matrix dr = {Magma_CSR}, drs = {Magma_CSR};
magma_d_matrix dP = {Magma_CSR}, dP1 = {Magma_CSR};
magma_d_matrix dG = {Magma_CSR};
magma_d_matrix dU = {Magma_CSR};
magma_d_matrix dM = {Magma_CSR};
magma_d_matrix df = {Magma_CSR};
magma_d_matrix dt = {Magma_CSR};
magma_d_matrix dc = {Magma_CSR};
magma_d_matrix dv = {Magma_CSR};
magma_d_matrix dbeta = {Magma_CSR}, hbeta = {Magma_CSR};
// chronometry
real_Double_t tempo1, tempo2;
// initial s space
// TODO: add option for 's' (shadow space number)
// Hack: uses '--restart' option as the shadow space number.
// This is not a good idea because the default value of restart option is used to detect
// if the user provided a custom restart. This means that if the default restart value
// is changed then the code will think it was the user (unless the default value is
// also updated in the 'if' statement below.
s = 1;
if ( solver_par->restart != 50 ) {
if ( solver_par->restart > A.num_cols ) {
s = A.num_cols;
} else {
s = solver_par->restart;
}
}
solver_par->restart = s;
// set max iterations
solver_par->maxiter = min( 2 * A.num_cols, solver_par->maxiter );
// check if matrix A is square
if ( A.num_rows != A.num_cols ) {
//printf("Matrix A is not square.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
// |b|
nrmb = magma_dnrm2( b.num_rows, b.dval, 1, queue );
if ( nrmb == 0.0 ) {
magma_dscal( x->num_rows, MAGMA_D_ZERO, x->dval, 1, queue );
info = MAGMA_SUCCESS;
goto cleanup;
}
// r = b - A x
CHECK( magma_dvinit( &dr, Magma_DEV, b.num_rows, 1, c_zero, queue ));
CHECK( magma_dresidualvec( A, b, *x, &dr, &nrmr, queue ));
// |r|
solver_par->init_res = nrmr;
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nrmr;
}
// check if initial is guess good enough
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// P = randn(n, s)
// P = ortho(P)
//---------------------------------------
// P = 0.0
CHECK( magma_dvinit( &dP, Magma_CPU, A.num_cols, s, c_zero, queue ));
// P = randn(n, s)
distr = 3; // 1 = unif (0,1), 2 = unif (-1,1), 3 = normal (0,1)
dof = dP.num_rows * dP.num_cols;
lapackf77_dlarnv( &distr, iseed, &dof, dP.val );
// transfer P to device
CHECK( magma_dmtransfer( dP, &dP1, Magma_CPU, Magma_DEV, queue ));
magma_dmfree( &dP, queue );
// P = ortho(P1)
if ( dP1.num_cols > 1 ) {
// P = magma_dqr(P1), QR factorization
CHECK( magma_dqr( dP1.num_rows, dP1.num_cols, dP1, dP1.ld, &dP, NULL, queue ));
} else {
// P = P1 / |P1|
nrm = magma_dnrm2( dof, dP1.dval, 1, queue );
nrm = 1.0 / nrm;
magma_dscal( dof, nrm, dP1.dval, 1, queue );
CHECK( magma_dmtransfer( dP1, &dP, Magma_DEV, Magma_DEV, queue ));
}
magma_dmfree( &dP1, queue );
//---------------------------------------
// allocate memory for the scalar products
CHECK( magma_dvinit( &hbeta, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_dvinit( &dbeta, Magma_DEV, s, 1, c_zero, queue ));
// smoothing enabled
if ( smoothing > 0 ) {
// set smoothing solution vector
CHECK( magma_dmtransfer( *x, &dxs, Magma_DEV, Magma_DEV, queue ));
// set smoothing residual vector
CHECK( magma_dmtransfer( dr, &drs, Magma_DEV, Magma_DEV, queue ));
}
// G(n,s) = 0
CHECK( magma_dvinit( &dG, Magma_DEV, A.num_cols, s, c_zero, queue ));
// U(n,s) = 0
CHECK( magma_dvinit( &dU, Magma_DEV, A.num_cols, s, c_zero, queue ));
// M(s,s) = I
CHECK( magma_dvinit( &dM, Magma_DEV, s, s, c_zero, queue ));
magmablas_dlaset( MagmaFull, s, s, c_zero, c_one, dM.dval, s, queue );
// f = 0
CHECK( magma_dvinit( &df, Magma_DEV, dP.num_cols, 1, c_zero, queue ));
// t = 0
CHECK( magma_dvinit( &dt, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
// c = 0
CHECK( magma_dvinit( &dc, Magma_DEV, dM.num_cols, 1, c_zero, queue ));
// v = 0
CHECK( magma_dvinit( &dv, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
//--------------START TIME---------------
// chronometry
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->timing[0] = 0.0;
}
om = MAGMA_D_ONE;
innerflag = 0;
// start iteration
do
{
solver_par->numiter++;
// new RHS for small systems
// f = P' r
magmablas_dgemv( MagmaConjTrans, dP.num_rows, dP.num_cols, c_one, dP.dval, dP.ld, dr.dval, 1, c_zero, df.dval, 1, queue );
// shadow space loop
for ( k = 0; k < s; ++k ) {
sk = s - k;
// f(k:s) = M(k:s,k:s) c(k:s)
magma_dcopyvector( sk, &df.dval[k], 1, &dc.dval[k], 1, queue );
magma_dtrsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, sk, &dM.dval[k*dM.ld+k], dM.ld, &dc.dval[k], 1, queue );
// v = r - G(:,k:s) c(k:s)
magma_dcopyvector( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
magmablas_dgemv( MagmaNoTrans, dG.num_rows, sk, c_n_one, &dG.dval[k*dG.ld], dG.ld, &dc.dval[k], 1, c_one, dv.dval, 1, queue );
// U(:,k) = om * v + U(:,k:s) c(k:s)
magmablas_dgemv( MagmaNoTrans, dU.num_rows, sk, c_one, &dU.dval[k*dU.ld], dU.ld, &dc.dval[k], 1, om, dv.dval, 1, queue );
magma_dcopyvector( dU.num_rows, dv.dval, 1, &dU.dval[k*dU.ld], 1, queue );
// G(:,k) = A U(:,k)
CHECK( magma_d_spmv( c_one, A, dv, c_zero, dv, queue ));
solver_par->spmv_count++;
magma_dcopyvector( dG.num_rows, dv.dval, 1, &dG.dval[k*dG.ld], 1, queue );
// bi-orthogonalize the new basis vectors
for ( i = 0; i < k; ++i ) {
// alpha = P(:,i)' G(:,k)
alpha = magma_ddot( dP.num_rows, &dP.dval[i*dP.ld], 1, &dG.dval[k*dG.ld], 1, queue );
// alpha = alpha / M(i,i)
magma_dgetvector( 1, &dM.dval[i*dM.ld+i], 1, &mkk, 1, queue );
alpha = alpha / mkk;
// G(:,k) = G(:,k) - alpha * G(:,i)
magma_daxpy( dG.num_rows, -alpha, &dG.dval[i*dG.ld], 1, &dG.dval[k*dG.ld], 1, queue );
// U(:,k) = U(:,k) - alpha * U(:,i)
magma_daxpy( dU.num_rows, -alpha, &dU.dval[i*dU.ld], 1, &dU.dval[k*dU.ld], 1, queue );
}
// new column of M = P'G, first k-1 entries are zero
// M(k:s,k) = P(:,k:s)' G(:,k)
magmablas_dgemv( MagmaConjTrans, dP.num_rows, sk, c_one, &dP.dval[k*dP.ld], dP.ld, &dG.dval[k*dG.ld], 1, c_zero, &dM.dval[k*dM.ld+k], 1, queue );
// check M(k,k) == 0
magma_dgetvector( 1, &dM.dval[k*dM.ld+k], 1, &mkk, 1, queue );
if ( MAGMA_D_EQUAL(mkk, MAGMA_D_ZERO) ) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// beta = f(k) / M(k,k)
magma_dgetvector( 1, &df.dval[k], 1, &fk, 1, queue );
hbeta.val[k] = fk / mkk;
// check for nan
if ( magma_d_isnan( hbeta.val[k] ) || magma_d_isinf( hbeta.val[k] )) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// r = r - beta * G(:,k)
magma_daxpy( dr.num_rows, -hbeta.val[k], &dG.dval[k*dG.ld], 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
nrmr = magma_dnrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// x = x + beta * U(:,k)
magma_daxpy( x->num_rows, hbeta.val[k], &dU.dval[k*dU.ld], 1, x->dval, 1, queue );
// smoothing operation
//---------------------------------------
// t = rs - r
magma_dcopyvector( drs.num_rows, drs.dval, 1, dt.dval, 1, queue );
magma_daxpy( dt.num_rows, c_n_one, dr.dval, 1, dt.dval, 1, queue );
// t't
// t'rs
tt = magma_ddot( dt.num_rows, dt.dval, 1, dt.dval, 1, queue );
tr = magma_ddot( dt.num_rows, dt.dval, 1, drs.dval, 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = tr / tt;
// rs = rs - gamma * (rs - r)
magma_daxpy( drs.num_rows, -gamma, dt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_dcopyvector( dxs.num_rows, dxs.dval, 1, dt.dval, 1, queue );
magma_daxpy( dt.num_rows, c_n_one, x->dval, 1, dt.dval, 1, queue );
magma_daxpy( dxs.num_rows, -gamma, dt.dval, 1, dxs.dval, 1, queue );
// |rs|
nrmr = magma_dnrm2( drs.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
s = k + 1; // for the x-update outside the loop
innerflag = 2;
info = MAGMA_SUCCESS;
break;
}
// non-last s iteration
if ( (k + 1) < s ) {
// f(k+1:s) = f(k+1:s) - beta * M(k+1:s,k)
magma_daxpy( sk-1, -hbeta.val[k], &dM.dval[k*dM.ld+(k+1)], 1, &df.dval[k+1], 1, queue );
}
}
// smoothing disabled
if ( smoothing <= 0 && innerflag != 1 ) {
// update solution approximation x
// x = x + U(:,1:s) * beta(1:s)
magma_dsetvector( s, hbeta.val, 1, dbeta.dval, 1, queue );
magmablas_dgemv( MagmaNoTrans, dU.num_rows, s, c_one, dU.dval, dU.ld, dbeta.dval, 1, c_one, x->dval, 1, queue );
}
// check convergence or iteration limit or invalid result of inner loop
if ( innerflag > 0 ) {
break;
}
// t = A v
// t = A r
CHECK( magma_d_spmv( c_one, A, dr, c_zero, dt, queue ));
solver_par->spmv_count++;
// computation of a new omega
//---------------------------------------
// |t|
nrmt = magma_dnrm2( dt.num_rows, dt.dval, 1, queue );
// t'r
tr = magma_ddot( dt.num_rows, dt.dval, 1, dr.dval, 1, queue );
// rho = abs(t' * r) / (|t| * |r|))
rho = MAGMA_D_ABS( MAGMA_D_REAL(tr) / (nrmt * nrmr) );
// om = (t' * r) / (|t| * |t|)
om = tr / (nrmt * nrmt);
if ( rho < angle ) {
om = (om * angle) / rho;
}
//---------------------------------------
if ( MAGMA_D_EQUAL(om, MAGMA_D_ZERO) ) {
info = MAGMA_DIVERGENCE;
break;
}
// update approximation vector
// x = x + om * v
// x = x + om * r
magma_daxpy( x->num_rows, om, dr.dval, 1, x->dval, 1, queue );
// update residual vector
// r = r - om * t
magma_daxpy( dr.num_rows, -om, dt.dval, 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// residual norm
nrmr = magma_dnrm2( b.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
magma_dcopyvector( drs.num_rows, drs.dval, 1, dt.dval, 1, queue );
magma_daxpy( dt.num_rows, c_n_one, dr.dval, 1, dt.dval, 1, queue );
// t't
// t'rs
tt = magma_ddot( dt.num_rows, dt.dval, 1, dt.dval, 1, queue );
tr = magma_ddot( dt.num_rows, dt.dval, 1, drs.dval, 1, queue );
// gamma = (t' * rs) / (|t| * |t|)
gamma = tr / tt;
// rs = rs - gamma * (rs - r)
magma_daxpy( drs.num_rows, -gamma, dt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_dcopyvector( dxs.num_rows, dxs.dval, 1, dt.dval, 1, queue );
magma_daxpy( dt.num_rows, c_n_one, x->dval, 1, dt.dval, 1, queue );
magma_daxpy( dxs.num_rows, -gamma, dt.dval, 1, dxs.dval, 1, queue );
// |rs|
nrmr = magma_dnrm2( b.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
break;
}
}
while ( solver_par->numiter + 1 <= solver_par->maxiter );
// smoothing enabled
if ( smoothing > 0 ) {
// x = xs
magma_dcopyvector( x->num_rows, dxs.dval, 1, x->dval, 1, queue );
// r = rs
magma_dcopyvector( dr.num_rows, drs.dval, 1, dr.dval, 1, queue );
}
// get last iteration timing
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t)tempo2 - tempo1;
//--------------STOP TIME----------------
// get final stats
solver_par->iter_res = nrmr;
CHECK( magma_dresidualvec( A, b, *x, &dr, &residual, queue ));
solver_par->final_res = residual;
// set solver conclusion
if ( info != MAGMA_SUCCESS && info != MAGMA_DIVERGENCE ) {
if ( solver_par->init_res > solver_par->final_res ) {
info = MAGMA_SLOW_CONVERGENCE;
}
}
cleanup:
// free resources
// smoothing enabled
if ( smoothing > 0 ) {
magma_dmfree( &dxs, queue );
magma_dmfree( &drs, queue );
}
magma_dmfree( &dr, queue );
magma_dmfree( &dP, queue );
magma_dmfree( &dP1, queue );
magma_dmfree( &dG, queue );
magma_dmfree( &dU, queue );
magma_dmfree( &dM, queue );
magma_dmfree( &df, queue );
magma_dmfree( &dt, queue );
magma_dmfree( &dc, queue );
magma_dmfree( &dv, queue );
magma_dmfree( &dbeta, queue );
magma_dmfree( &hbeta, queue );
solver_par->info = info;
return info;
/* magma_didr */
}
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;
}
