magma_int_t
magma_scsrget_gpu(
magma_s_matrix A,
magma_int_t *m,
magma_int_t *n,
magmaIndex_ptr *row,
magmaIndex_ptr *col,
magmaFloat_ptr *val,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_s_matrix A_DEV={Magma_CSR}, A_CSR={Magma_CSR};
if ( A.memory_location == Magma_DEV && A.storage_type == Magma_CSR ) {
*m = A.num_rows;
*n = A.num_cols;
*val = A.dval;
*col = A.dcol;
*row = A.drow;
} else {
CHECK( magma_smconvert( A, &A_CSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_smtransfer( A_CSR, &A_DEV, A.memory_location, Magma_DEV, queue ));
magma_scsrget_gpu( A_DEV, m, n, row, col, val, queue );
}
cleanup:
magma_smfree( &A_CSR, queue );
magma_smfree( &A_DEV, queue );
return info;
}
extern "C" magma_int_t
magma_svtranspose(
magma_s_matrix x,
magma_s_matrix *y,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t m = x.num_rows;
magma_int_t n = x.num_cols;
// set queue for old dense routines
magma_queue_t orig_queue=NULL;
magmablasGetKernelStream( &orig_queue );
magma_s_matrix x_d={Magma_CSR}, y_d={Magma_CSR};
if ( x.memory_location == Magma_DEV ) {
CHECK( magma_svinit( y, Magma_DEV, x.num_rows,x.num_cols, MAGMA_S_ZERO, queue ));
y->num_rows = x.num_rows;
y->num_cols = x.num_cols;
y->storage_type = x.storage_type;
if ( x.major == MagmaColMajor) {
y->major = MagmaRowMajor;
magmablas_stranspose( m, n, x.val, m, y->val, n );
}
else {
y->major = MagmaColMajor;
magmablas_stranspose( n, m, x.val, n, y->val, m );
}
} else {
CHECK( magma_smtransfer( x, &x_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_svtranspose( x_d, &y_d, queue ));
CHECK( magma_smtransfer( y_d, y, Magma_DEV, Magma_CPU, queue ));
}
cleanup:
if( info != 0 ){
magma_smfree( y, queue );
}
magma_smfree( &x_d, queue );
magma_smfree( &y_d, queue );
magmablasSetKernelStream( orig_queue );
return info;
}
extern "C" magma_int_t
magma_smshrink(
magma_s_matrix A,
magma_s_matrix *B,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_s_matrix hA={Magma_CSR}, hACSR={Magma_CSR}, hB={Magma_CSR}, hBCSR={Magma_CSR};
if( A.num_rows<=A.num_cols){
if( A.memory_location == Magma_CPU && A.storage_type == Magma_CSR ){
CHECK( magma_smconvert( A, B, Magma_CSR, Magma_CSR, queue ));
for(magma_int_t i=0; i<A.nnz; i++){
if( B->col[i] >= A.num_rows ){
B->val[i] = MAGMA_S_ZERO;
}
}
CHECK( magma_smcsrcompressor( B, queue ) );
B->num_cols = B->num_rows;
} else {
CHECK( magma_smtransfer( A, &hA, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smconvert( hA, &hACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_smshrink( hACSR, &hBCSR, queue ));
CHECK( magma_smconvert( hBCSR, &hB, Magma_CSR, A.storage_type, queue ));
CHECK( magma_smtransfer( hB, B, Magma_CPU, A.memory_location, queue ));
}
} else {
printf("%% error: A has too many rows: m > n.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
cleanup:
magma_smfree( &hA, queue );
magma_smfree( &hB, queue );
magma_smfree( &hACSR, queue );
magma_smfree( &hBCSR, queue );
return info;
}
extern "C" magma_int_t
magma_sjacobisetup_vector(
magma_s_matrix b, magma_s_matrix d,
magma_s_matrix *c,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_s_matrix diag={Magma_CSR}, c_t={Magma_CSR}, b_h={Magma_CSR}, tmp={Magma_CSR};
if ( b.memory_location == Magma_CPU ) {
CHECK( magma_svinit( &c_t, Magma_CPU, b.num_rows, b.num_cols, MAGMA_S_ZERO, queue ));
CHECK( magma_smtransfer( b, &b_h, b.memory_location, Magma_CPU, queue ));
CHECK( magma_smtransfer( d, &diag, b.memory_location, Magma_CPU, queue ));
for( magma_int_t rowindex=0; rowindex<b.num_rows; rowindex++ ) {
c_t.val[rowindex] = b_h.val[rowindex] / diag.val[rowindex];
}
CHECK( magma_smtransfer( c_t, c, Magma_CPU, b.memory_location, queue ));
}
else if ( b.memory_location == Magma_DEV ) {
// fill vector
CHECK( magma_svinit( &tmp, Magma_DEV, b.num_rows, b.num_cols, MAGMA_S_ZERO, queue ));
CHECK( magma_sjacobisetup_vector_gpu(
b.num_rows, b, d, *c, &tmp, queue ));
goto cleanup;
}
cleanup:
magma_smfree( &tmp, queue );
magma_smfree( &diag, queue );
magma_smfree( &c_t, queue );
magma_smfree( &b_h, queue );
return info;
}
magma_int_t
magma_svget(
magma_s_matrix v,
magma_int_t *m, magma_int_t *n,
float **val,
magma_queue_t queue )
{
magma_s_matrix v_CPU={Magma_CSR};
magma_int_t info =0;
if ( v.memory_location == Magma_CPU ) {
*m = v.num_rows;
*n = v.num_cols;
*val = v.val;
} else {
CHECK( magma_smtransfer( v, &v_CPU, v.memory_location, Magma_CPU, queue ));
CHECK( magma_svget( v_CPU, m, n, val, queue ));
}
cleanup:
magma_smfree( &v_CPU, queue );
return info;
}
magma_int_t
magma_svget_dev(
magma_s_matrix v,
magma_int_t *m, magma_int_t *n,
magmaFloat_ptr *val,
magma_queue_t queue )
{
magma_int_t info =0;
magma_s_matrix v_DEV={Magma_CSR};
if ( v.memory_location == Magma_DEV ) {
*m = v.num_rows;
*n = v.num_cols;
*val = v.dval;
} else {
CHECK( magma_smtransfer( v, &v_DEV, v.memory_location, Magma_DEV, queue ));
CHECK( magma_svget_dev( v_DEV, m, n, val, queue ));
}
cleanup:
magma_smfree( &v_DEV, queue );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_s_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
CHECK( magma_sparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
CHECK( magma_ssolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue ));
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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
// for the eigensolver case
zopts.solver_par.ev_length = A.num_cols;
CHECK( magma_seigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_smscale( &A, zopts.scaling, queue ));
// preconditioner
if ( zopts.solver_par.solver != Magma_ITERREF ) {
CHECK( magma_s_precondsetup( A, b, &zopts.solver_par, &zopts.precond_par, queue ) );
}
CHECK( magma_smconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
printf( "\n%% matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
printf("matrixinfo = [ \n");
printf("%% size (m x n) || nonzeros (nnz) || nnz/m || stored nnz\n");
printf("%%======================================================================"
"======%%\n");
printf(" %8d %8d %10d %4d %10d\n",
int(B.num_rows), int(B.num_cols), int(B.true_nnz), int(B.true_nnz/B.num_rows), int(B.nnz) );
printf("%%======================================================================"
"======%%\n");
printf("];\n");
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_rows, 1, one, queue ));
//magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_s_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_smfree(&x, queue );
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_s_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ) {
printf("%%error: solver returned: %s (%d).\n",
magma_strerror( info ), int(info) );
}
printf("data = [\n");
magma_ssolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
printf("];\n\n");
printf("precond_info = [\n");
printf("%% setup runtime\n");
printf(" %.6f %.6f\n",
zopts.precond_par.setuptime, zopts.precond_par.runtime );
printf("];\n\n");
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_ssolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
/* Initialize */
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( &queue );
magmablasSetKernelStream( queue );
magma_int_t j, n=1000000, FLOPS;
float one = MAGMA_S_MAKE( 1.0, 0.0 );
float two = MAGMA_S_MAKE( 2.0, 0.0 );
magma_s_matrix a={Magma_CSR}, ad={Magma_CSR}, bd={Magma_CSR}, cd={Magma_CSR};
CHECK( magma_svinit( &a, Magma_CPU, n, 1, one, queue ));
CHECK( magma_svinit( &bd, Magma_DEV, n, 1, two, queue ));
CHECK( magma_svinit( &cd, Magma_DEV, n, 1, one, queue ));
CHECK( magma_smtransfer( a, &ad, Magma_CPU, Magma_DEV, queue ));
real_Double_t start, end, res;
FLOPS = 2*n;
start = magma_sync_wtime( queue );
for (j=0; j<100; j++)
res = magma_snrm2(n, ad.dval, 1);
end = magma_sync_wtime( queue );
printf( " > MAGMA nrm2: %.2e seconds %.2e GFLOP/s\n",
(end-start)/100, FLOPS*100/1e9/(end-start) );
FLOPS = n;
start = magma_sync_wtime( queue );
for (j=0; j<100; j++)
magma_sscal( n, two, ad.dval, 1 );
end = magma_sync_wtime( queue );
printf( " > MAGMA scal: %.2e seconds %.2e GFLOP/s\n",
(end-start)/100, FLOPS*100/1e9/(end-start) );
FLOPS = 2*n;
start = magma_sync_wtime( queue );
for (j=0; j<100; j++)
magma_saxpy( n, one, ad.dval, 1, bd.dval, 1 );
end = magma_sync_wtime( queue );
printf( " > MAGMA axpy: %.2e seconds %.2e GFLOP/s\n",
(end-start)/100, FLOPS*100/1e9/(end-start) );
FLOPS = n;
start = magma_sync_wtime( queue );
for (j=0; j<100; j++)
magma_scopy( n, bd.dval, 1, ad.dval, 1 );
end = magma_sync_wtime( queue );
printf( " > MAGMA copy: %.2e seconds %.2e GFLOP/s\n",
(end-start)/100, FLOPS*100/1e9/(end-start) );
FLOPS = 2*n;
start = magma_sync_wtime( queue );
for (j=0; j<100; j++)
res = MAGMA_S_REAL( magma_sdot(n, ad.dval, 1, bd.dval, 1) );
end = magma_sync_wtime( queue );
printf( " > MAGMA dotc: %.2e seconds %.2e GFLOP/s\n",
(end-start)/100, FLOPS*100/1e9/(end-start) );
printf("# tester BLAS: ok\n");
magma_smfree( &a, queue);
magma_smfree(&ad, queue);
magma_smfree(&bd, queue);
magma_smfree(&cd, queue);
cleanup:
magma_smfree( &a, queue);
magma_smfree(&ad, queue);
magma_smfree(&bd, queue);
magma_smfree(&cd, queue);
magmablasSetKernelStream( NULL );
magma_queue_destroy( queue );
magma_finalize();
return info;
}
extern "C" magma_int_t
magma_spidr(
magma_s_matrix A, magma_s_matrix b, magma_s_matrix *x,
magma_s_solver_par *solver_par,
magma_s_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_PIDR;
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 float c_zero = MAGMA_S_ZERO;
const float c_one = MAGMA_S_ONE;
const float c_n_one = MAGMA_S_NEG_ONE;
// internal user parameters
const magma_int_t smoothing = 1; // 0 = disable, 1 = enable
const float 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;
float residual;
float nrm;
float nrmb;
float nrmr;
float nrmt;
float rho;
float om;
float tt;
float tr;
float gamma;
float alpha;
float mkk;
float fk;
// matrices and vectors
magma_s_matrix dxs = {Magma_CSR};
magma_s_matrix dr = {Magma_CSR}, drs = {Magma_CSR};
magma_s_matrix dP = {Magma_CSR}, dP1 = {Magma_CSR};
magma_s_matrix dG = {Magma_CSR};
magma_s_matrix dU = {Magma_CSR};
magma_s_matrix dM = {Magma_CSR};
magma_s_matrix df = {Magma_CSR};
magma_s_matrix dt = {Magma_CSR};
magma_s_matrix dc = {Magma_CSR};
magma_s_matrix dv = {Magma_CSR};
magma_s_matrix dbeta = {Magma_CSR}, hbeta = {Magma_CSR};
magma_s_matrix dlu = {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_snrm2( b.num_rows, b.dval, 1, queue );
if ( nrmb == 0.0 ) {
magma_sscal( x->num_rows, MAGMA_S_ZERO, x->dval, 1, queue );
info = MAGMA_SUCCESS;
goto cleanup;
}
// r = b - A x
CHECK( magma_svinit( &dr, Magma_DEV, b.num_rows, 1, c_zero, queue ));
CHECK( magma_sresidualvec( 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_svinit( &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_slarnv( &distr, iseed, &dof, dP.val );
// transfer P to device
CHECK( magma_smtransfer( dP, &dP1, Magma_CPU, Magma_DEV, queue ));
magma_smfree( &dP, queue );
// P = ortho(P1)
if ( dP1.num_cols > 1 ) {
// P = magma_sqr(P1), QR factorization
CHECK( magma_sqr( dP1.num_rows, dP1.num_cols, dP1, dP1.ld, &dP, NULL, queue ));
} else {
// P = P1 / |P1|
nrm = magma_snrm2( dof, dP1.dval, 1, queue );
nrm = 1.0 / nrm;
magma_sscal( dof, nrm, dP1.dval, 1, queue );
CHECK( magma_smtransfer( dP1, &dP, Magma_DEV, Magma_DEV, queue ));
}
magma_smfree( &dP1, queue );
//---------------------------------------
// allocate memory for the scalar products
CHECK( magma_svinit( &hbeta, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_svinit( &dbeta, Magma_DEV, s, 1, c_zero, queue ));
// smoothing enabled
if ( smoothing > 0 ) {
// set smoothing solution vector
CHECK( magma_smtransfer( *x, &dxs, Magma_DEV, Magma_DEV, queue ));
// set smoothing residual vector
CHECK( magma_smtransfer( dr, &drs, Magma_DEV, Magma_DEV, queue ));
}
// G(n,s) = 0
CHECK( magma_svinit( &dG, Magma_DEV, A.num_cols, s, c_zero, queue ));
// U(n,s) = 0
CHECK( magma_svinit( &dU, Magma_DEV, A.num_cols, s, c_zero, queue ));
// M(s,s) = I
CHECK( magma_svinit( &dM, Magma_DEV, s, s, c_zero, queue ));
magmablas_slaset( MagmaFull, s, s, c_zero, c_one, dM.dval, s, queue );
// f = 0
CHECK( magma_svinit( &df, Magma_DEV, dP.num_cols, 1, c_zero, queue ));
// t = 0
CHECK( magma_svinit( &dt, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
// c = 0
CHECK( magma_svinit( &dc, Magma_DEV, dM.num_cols, 1, c_zero, queue ));
// v = 0
CHECK( magma_svinit( &dv, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
// lu = 0
CHECK( magma_svinit( &dlu, Magma_DEV, A.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_S_ONE;
innerflag = 0;
// start iteration
do
{
solver_par->numiter++;
// new RHS for small systems
// f = P' r
magmablas_sgemv( 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_scopyvector( sk, &df.dval[k], 1, &dc.dval[k], 1, queue );
magma_strsv( 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_scopyvector( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
magmablas_sgemv( 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 );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_s_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_s_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// U(:,k) = om * v + U(:,k:s) c(k:s)
magmablas_sgemv( MagmaNoTrans, dU.num_rows, sk, c_one, &dU.dval[k*dU.ld], dU.ld, &dc.dval[k], 1, om, dv.dval, 1, queue );
magma_scopyvector( dU.num_rows, dv.dval, 1, &dU.dval[k*dU.ld], 1, queue );
// G(:,k) = A U(:,k)
CHECK( magma_s_spmv( c_one, A, dv, c_zero, dv, queue ));
solver_par->spmv_count++;
magma_scopyvector( 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_sdot( dP.num_rows, &dP.dval[i*dP.ld], 1, &dG.dval[k*dG.ld], 1, queue );
// alpha = alpha / M(i,i)
magma_sgetvector( 1, &dM.dval[i*dM.ld+i], 1, &mkk, 1, queue );
alpha = alpha / mkk;
// G(:,k) = G(:,k) - alpha * G(:,i)
magma_saxpy( 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_saxpy( 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_sgemv( 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_sgetvector( 1, &dM.dval[k*dM.ld+k], 1, &mkk, 1, queue );
if ( MAGMA_S_EQUAL(mkk, MAGMA_S_ZERO) ) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// beta = f(k) / M(k,k)
magma_sgetvector( 1, &df.dval[k], 1, &fk, 1, queue );
hbeta.val[k] = fk / mkk;
// check for nan
if ( magma_s_isnan( hbeta.val[k] ) || magma_s_isinf( hbeta.val[k] )) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// r = r - beta * G(:,k)
magma_saxpy( dr.num_rows, -hbeta.val[k], &dG.dval[k*dG.ld], 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
nrmr = magma_snrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// x = x + beta * U(:,k)
magma_saxpy( x->num_rows, hbeta.val[k], &dU.dval[k*dU.ld], 1, x->dval, 1, queue );
// smoothing operation
//---------------------------------------
// t = rs - r
magma_scopyvector( drs.num_rows, drs.dval, 1, dt.dval, 1, queue );
magma_saxpy( dt.num_rows, c_n_one, dr.dval, 1, dt.dval, 1, queue );
// t't
// t'rs
tt = magma_sdot( dt.num_rows, dt.dval, 1, dt.dval, 1, queue );
tr = magma_sdot( dt.num_rows, dt.dval, 1, drs.dval, 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = tr / tt;
// rs = rs - gamma * (rs - r)
magma_saxpy( drs.num_rows, -gamma, dt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_scopyvector( dxs.num_rows, dxs.dval, 1, dt.dval, 1, queue );
magma_saxpy( dt.num_rows, c_n_one, x->dval, 1, dt.dval, 1, queue );
magma_saxpy( dxs.num_rows, -gamma, dt.dval, 1, dxs.dval, 1, queue );
// |rs|
nrmr = magma_snrm2( 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_saxpy( 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_ssetvector( s, hbeta.val, 1, dbeta.dval, 1, queue );
magmablas_sgemv( 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;
}
// v = r
magma_scopyvector( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_s_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_s_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// t = A v
CHECK( magma_s_spmv( c_one, A, dv, c_zero, dt, queue ));
solver_par->spmv_count++;
// computation of a new omega
//---------------------------------------
// |t|
nrmt = magma_snrm2( dt.num_rows, dt.dval, 1, queue );
// t'r
tr = magma_sdot( dt.num_rows, dt.dval, 1, dr.dval, 1, queue );
// rho = abs(t' * r) / (|t| * |r|))
rho = MAGMA_D_ABS( MAGMA_S_REAL(tr) / (nrmt * nrmr) );
// om = (t' * r) / (|t| * |t|)
om = tr / (nrmt * nrmt);
if ( rho < angle ) {
om = (om * angle) / rho;
}
//---------------------------------------
if ( MAGMA_S_EQUAL(om, MAGMA_S_ZERO) ) {
info = MAGMA_DIVERGENCE;
break;
}
// update approximation vector
// x = x + om * v
magma_saxpy( x->num_rows, om, dv.dval, 1, x->dval, 1, queue );
// update residual vector
// r = r - om * t
magma_saxpy( dr.num_rows, -om, dt.dval, 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// residual norm
nrmr = magma_snrm2( b.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
magma_scopyvector( drs.num_rows, drs.dval, 1, dt.dval, 1, queue );
magma_saxpy( dt.num_rows, c_n_one, dr.dval, 1, dt.dval, 1, queue );
// t't
// t'rs
tt = magma_sdot( dt.num_rows, dt.dval, 1, dt.dval, 1, queue );
tr = magma_sdot( dt.num_rows, dt.dval, 1, drs.dval, 1, queue );
// gamma = (t' * rs) / (|t| * |t|)
gamma = tr / tt;
// rs = rs - gamma * (rs - r)
magma_saxpy( drs.num_rows, -gamma, dt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_scopyvector( dxs.num_rows, dxs.dval, 1, dt.dval, 1, queue );
magma_saxpy( dt.num_rows, c_n_one, x->dval, 1, dt.dval, 1, queue );
magma_saxpy( dxs.num_rows, -gamma, dt.dval, 1, dxs.dval, 1, queue );
// |rs|
nrmr = magma_snrm2( 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_scopyvector( x->num_rows, dxs.dval, 1, x->dval, 1, queue );
// r = rs
magma_scopyvector( 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_sresidualvec( 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_smfree( &dxs, queue );
magma_smfree( &drs, queue );
}
magma_smfree( &dr, queue );
magma_smfree( &dP, queue );
magma_smfree( &dP1, queue );
magma_smfree( &dG, queue );
magma_smfree( &dU, queue );
magma_smfree( &dM, queue );
magma_smfree( &df, queue );
magma_smfree( &dt, queue );
magma_smfree( &dc, queue );
magma_smfree( &dv, queue );
magma_smfree(&dlu, queue);
magma_smfree( &dbeta, queue );
magma_smfree( &hbeta, queue );
solver_par->info = info;
return info;
/* magma_spidr */
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_CHECK( magma_init() );
magma_print_environment();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magma_s_matrix hA={Magma_CSR}, hA_SELLP={Magma_CSR},
dA={Magma_CSR}, dA_SELLP={Magma_CSR};
magma_s_matrix hx={Magma_CSR}, hy={Magma_CSR}, dx={Magma_CSR},
dy={Magma_CSR}, hrefvec={Magma_CSR}, hcheck={Magma_CSR};
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
#ifdef MAGMA_WITH_MKL
magma_int_t *pntre=NULL;
#endif
cusparseHandle_t cusparseHandle = NULL;
cusparseMatDescr_t descr = NULL;
float c_one = MAGMA_S_MAKE(1.0, 0.0);
float c_zero = MAGMA_S_MAKE(0.0, 0.0);
float accuracy = 1e-10;
#define PRECISION_s
#if defined(PRECISION_c)
accuracy = 1e-4;
#endif
#if defined(PRECISION_s)
accuracy = 1e-4;
#endif
magma_int_t i, j;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--blocksize", argv[i]) == 0 ) {
hA_SELLP.blocksize = atoi( argv[++i] );
} else if ( strcmp("--alignment", argv[i]) == 0 ) {
hA_SELLP.alignment = atoi( argv[++i] );
} else
break;
}
printf("\n# usage: ./run_sspmm"
" [ --blocksize %lld --alignment %lld (for SELLP) ] matrices\n\n",
(long long) hA_SELLP.blocksize, (long long) hA_SELLP.alignment );
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
TESTING_CHECK( magma_sm_5stencil( laplace_size, &hA, queue ));
} else { // file-matrix test
TESTING_CHECK( magma_s_csr_mtx( &hA, argv[i], queue ));
}
printf("%% matrix info: %lld-by-%lld with %lld nonzeros\n",
(long long) hA.num_rows, (long long) hA.num_cols, (long long) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
// m - number of rows for the sparse matrix
// n - number of vectors to be multiplied in the SpMM product
magma_int_t m, n;
m = hA.num_rows;
n = 48;
// init CPU vectors
TESTING_CHECK( magma_svinit( &hx, Magma_CPU, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &hy, Magma_CPU, m, n, c_zero, queue ));
// init DEV vectors
TESTING_CHECK( magma_svinit( &dx, Magma_DEV, m, n, c_one, queue ));
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, m, n, c_zero, queue ));
// calling MKL with CSR
#ifdef MAGMA_WITH_MKL
TESTING_CHECK( magma_imalloc_cpu( &pntre, m + 1 ) );
pntre[0] = 0;
for (j=0; j < m; j++ ) {
pntre[j] = hA.row[j+1];
}
MKL_INT num_rows = hA.num_rows;
MKL_INT num_cols = hA.num_cols;
MKL_INT nnz = hA.nnz;
MKL_INT num_vecs = n;
MKL_INT *col;
TESTING_CHECK( magma_malloc_cpu( (void**) &col, nnz * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_CHECK( magma_malloc_cpu( (void**) &row, num_rows * sizeof(MKL_INT) ));
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
// === Call MKL with consecutive SpMVs, using mkl_scsrmv ===
// warmp up
mkl_scsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
start = magma_wtime();
for (j=0; j < 10; j++ ) {
mkl_scsrmv( "N", &num_rows, &num_cols,
MKL_ADDR(&c_one), "GFNC", MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val),
MKL_ADDR(&c_zero), MKL_ADDR(hy.val) );
}
end = magma_wtime();
printf( "\n > MKL SpMVs : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
// === Call MKL with blocked SpMVs, using mkl_scsrmm ===
char transa = 'n';
MKL_INT ldb = n, ldc=n;
char matdescra[6] = {'g', 'l', 'n', 'c', 'x', 'x'};
// warm up
mkl_scsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
start = magma_wtime();
for (j=0; j < 10; j++ ) {
mkl_scsrmm( &transa, &num_rows, &num_vecs, &num_cols, MKL_ADDR(&c_one), matdescra,
MKL_ADDR(hA.val), col, row, pntre,
MKL_ADDR(hx.val), &ldb,
MKL_ADDR(&c_zero),
MKL_ADDR(hy.val), &ldc );
}
end = magma_wtime();
printf( "\n > MKL SpMM : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
magma_free_cpu( row );
magma_free_cpu( col );
row = NULL;
col = NULL;
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue ));
// SpMV on GPU (CSR)
start = magma_sync_wtime( queue );
for (j=0; j < 10; j++) {
TESTING_CHECK( magma_s_spmv( c_one, dA, dx, c_zero, dy, queue ));
}
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard CSR).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue ));
magma_smfree(&dA, queue );
// convert to SELLP and copy to GPU
TESTING_CHECK( magma_smconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue ));
TESTING_CHECK( magma_smtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue ));
magma_smfree(&hA_SELLP, queue );
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j < 10; j++) {
TESTING_CHECK( magma_s_spmv( c_one, dA_SELLP, dx, c_zero, dy, queue ));
}
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (SELLP).\n",
(end-start)/10, FLOPS*10.*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res=res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_REAL(hrefvec.val[k]);
}
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm SELL-P: ok\n");
else
printf("%% tester spmm SELL-P: failed\n");
magma_smfree( &hcheck, queue );
magma_smfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
magma_smfree( &dy, queue );
TESTING_CHECK( magma_svinit( &dy, Magma_DEV, dx.num_rows, dx.num_cols, c_zero, queue ));
//#ifdef PRECISION_d
start = magma_sync_wtime( queue );
TESTING_CHECK( cusparseCreate( &cusparseHandle ));
TESTING_CHECK( cusparseSetStream( cusparseHandle, magma_queue_get_cuda_stream(queue) ));
TESTING_CHECK( cusparseCreateMatDescr( &descr ));
TESTING_CHECK( cusparseSetMatType( descr, CUSPARSE_MATRIX_TYPE_GENERAL ));
TESTING_CHECK( cusparseSetMatIndexBase( descr, CUSPARSE_INDEX_BASE_ZERO ));
float alpha = c_one;
float beta = c_zero;
// copy matrix to GPU
TESTING_CHECK( magma_smtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue) );
for (j=0; j < 10; j++) {
cusparseScsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
dA.num_rows, n, dA.num_cols, dA.nnz,
&alpha, descr, dA.dval, dA.drow, dA.dcol,
dx.dval, dA.num_cols, &beta, dy.dval, dA.num_cols);
}
end = magma_sync_wtime( queue );
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10*n/(end-start) );
TESTING_CHECK( magma_smtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue ));
res = 0.0;
for(magma_int_t k=0; k < hA.num_rows; k++ ) {
res = res + MAGMA_S_REAL(hcheck.val[k]) - MAGMA_S_REAL(hrefvec.val[k]);
}
printf("%% |x-y|_F = %8.2e\n", res);
if ( res < accuracy )
printf("%% tester spmm cuSPARSE: ok\n");
else
printf("%% tester spmm cuSPARSE: failed\n");
magma_smfree( &hcheck, queue );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
descr = NULL;
cusparseHandle = NULL;
//#endif
printf("\n\n");
// free CPU memory
magma_smfree( &hA, queue );
magma_smfree( &hx, queue );
magma_smfree( &hy, queue );
magma_smfree( &hrefvec, queue );
// free GPU memory
magma_smfree( &dx, queue );
magma_smfree( &dy, queue );
magma_smfree( &dA, queue);
#ifdef MAGMA_WITH_MKL
magma_free_cpu( pntre );
#endif
i++;
}
magma_queue_destroy( queue );
TESTING_CHECK( magma_finalize() );
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B_d={Magma_CSR};
magma_s_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
CHECK( magma_sparse_opts( argc, argv, &zopts, &i, queue ));
B.blocksize = zopts.blocksize;
B.alignment = zopts.alignment;
if ( zopts.solver_par.solver != Magma_PCG &&
zopts.solver_par.solver != Magma_PGMRES &&
zopts.solver_par.solver != Magma_PBICGSTAB &&
zopts.solver_par.solver != Magma_ITERREF &&
zopts.solver_par.solver != Magma_LOBPCG )
zopts.precond_par.solver = Magma_NONE;
CHECK( magma_ssolverinfo_init( &zopts.solver_par, &zopts.precond_par, queue ));
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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// for the eigensolver case
zopts.solver_par.ev_length = A.num_rows;
CHECK( magma_seigensolverinfo_init( &zopts.solver_par, queue ));
// scale matrix
CHECK( magma_smscale( &A, zopts.scaling, queue ));
CHECK( magma_smconvert( A, &B, Magma_CSR, zopts.output_format, queue ));
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_cols, 1, one, queue ));
//magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue );
//magma_s_spmv( one, B_d, x, zero, b, queue ); // b = A x
//magma_smfree(&x, queue );
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, zero, queue ));
info = magma_s_solver( B_d, b, &x, &zopts, queue );
if( info != 0 ){
printf("error: solver returned: %s (%d).\n",
magma_strerror( info ), info );
}
magma_ssolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&B_d, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_ssolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
extern "C" magma_int_t
magma_sbicg(
magma_s_matrix A, magma_s_matrix b, magma_s_matrix *x,
magma_s_solver_par *solver_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_BICG;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// some useful variables
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t dofs = A.num_rows * b.num_cols;
// workspace
magma_s_matrix r={Magma_CSR}, rt={Magma_CSR}, p={Magma_CSR}, pt={Magma_CSR},
z={Magma_CSR}, zt={Magma_CSR}, q={Magma_CSR}, y={Magma_CSR},
yt={Magma_CSR}, qt={Magma_CSR};
// need to transpose the matrix
magma_s_matrix AT={Magma_CSR}, Ah1={Magma_CSR}, Ah2={Magma_CSR};
CHECK( magma_svinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &rt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &pt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &qt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &y, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &yt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &zt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver variables
float alpha, rho, beta, rho_new, ptq;
float res, nomb, nom0, r0;
// transpose the matrix
magma_smtransfer( A, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_smconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_smfree(&Ah1, queue );
magma_smtransposeconjugate( Ah2, &Ah1, queue );
magma_smfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_smconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_smfree(&Ah1, queue );
magma_smtransfer( Ah2, &AT, Magma_CPU, Magma_DEV, queue );
magma_smfree(&Ah2, queue );
// solver setup
CHECK( magma_sresidualvec( A, b, *x, &r, &nom0, queue));
res = nom0;
solver_par->init_res = nom0;
magma_scopy( dofs, r.dval, 1, rt.dval, 1, queue ); // rr = r
rho_new = magma_sdot( dofs, rt.dval, 1, r.dval, 1, queue ); // rho=<rr,r>
rho = alpha = MAGMA_S_MAKE( 1.0, 0. );
nomb = magma_snrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
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] = nom0;
solver_par->timing[0] = 0.0;
}
if ( nom0 < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
magma_scopy( dofs, r.dval, 1 , y.dval, 1, queue ); // y=r
magma_scopy( dofs, y.dval, 1 , z.dval, 1, queue ); // z=y
magma_scopy( dofs, rt.dval, 1 , yt.dval, 1, queue ); // yt=rt
magma_scopy( dofs, yt.dval, 1 , zt.dval, 1, queue ); // zt=yt
rho= rho_new;
rho_new = magma_sdot( dofs, rt.dval, 1, z.dval, 1, queue ); // rho=<rt,z>
if( magma_s_isnan_inf( rho_new ) ){
info = MAGMA_DIVERGENCE;
break;
}
if( solver_par->numiter==1 ){
magma_scopy( dofs, z.dval, 1 , p.dval, 1, queue ); // yt=rt
magma_scopy( dofs, zt.dval, 1 , pt.dval, 1, queue ); // zt=yt
} else {
beta = rho_new/rho;
magma_sscal( dofs, beta, p.dval, 1, queue ); // p = beta*p
magma_saxpy( dofs, c_one , z.dval, 1 , p.dval, 1, queue ); // p = z+beta*p
magma_sscal( dofs, MAGMA_S_CONJ(beta), pt.dval, 1, queue ); // pt = beta*pt
magma_saxpy( dofs, c_one , zt.dval, 1 , pt.dval, 1, queue ); // pt = zt+beta*pt
}
CHECK( magma_s_spmv( c_one, A, p, c_zero, q, queue )); // v = Ap
CHECK( magma_s_spmv( c_one, AT, pt, c_zero, qt, queue )); // v = Ap
solver_par->spmv_count++;
solver_par->spmv_count++;
ptq = magma_sdot( dofs, pt.dval, 1, q.dval, 1, queue );
alpha = rho_new /ptq;
magma_saxpy( dofs, alpha, p.dval, 1 , x->dval, 1, queue ); // x=x+alpha*p
magma_saxpy( dofs, c_neg_one * alpha, q.dval, 1 , r.dval, 1, queue ); // r=r+alpha*q
magma_saxpy( dofs, c_neg_one * MAGMA_S_CONJ(alpha), qt.dval, 1 , rt.dval, 1, queue ); // r=r+alpha*q
res = magma_snrm2( dofs, r.dval, 1, queue );
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) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
CHECK( magma_sresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_smfree(&r, queue );
magma_smfree(&rt, queue );
magma_smfree(&p, queue );
magma_smfree(&pt, queue );
magma_smfree(&q, queue );
magma_smfree(&qt, queue );
magma_smfree(&y, queue );
magma_smfree(&yt, queue );
magma_smfree(&z, queue );
magma_smfree(&zt, queue );
magma_smfree(&AT, queue );
magma_smfree(&Ah1, queue );
magma_smfree(&Ah2, queue );
solver_par->info = info;
return info;
} /* magma_sbicg */
extern "C" magma_int_t
magma_smtransposeconjugate(
magma_s_matrix A,
magma_s_matrix *B,
magma_queue_t queue )
{
// for symmetric matrices: convert to csc using cusparse
magma_int_t info = 0;
cusparseHandle_t handle=NULL;
cusparseMatDescr_t descrA=NULL;
cusparseMatDescr_t descrB=NULL;
magma_s_matrix ACSR={Magma_CSR}, BCSR={Magma_CSR};
magma_s_matrix A_d={Magma_CSR}, B_d={Magma_CSR};
if( A.storage_type == Magma_CSR && A.memory_location == Magma_DEV ) {
// fill in information for B
B->storage_type = A.storage_type;
B->diagorder_type = A.diagorder_type;
B->memory_location = Magma_DEV;
B->num_rows = A.num_cols; // transposed
B->num_cols = A.num_rows; // transposed
B->nnz = A.nnz;
B->true_nnz = A.true_nnz;
if ( A.fill_mode == MagmaFull ) {
B->fill_mode = MagmaFull;
}
else if ( A.fill_mode == MagmaLower ) {
B->fill_mode = MagmaUpper;
}
else if ( A.fill_mode == MagmaUpper ) {
B->fill_mode = MagmaLower;
}
B->dval = NULL;
B->drow = NULL;
B->dcol = NULL;
// memory allocation
CHECK( magma_smalloc( &B->dval, B->nnz ));
CHECK( magma_index_malloc( &B->drow, B->num_rows + 1 ));
CHECK( magma_index_malloc( &B->dcol, B->nnz ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &handle ));
CHECK_CUSPARSE( cusparseSetStream( handle, queue->cuda_stream() ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrA ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrB ));
CHECK_CUSPARSE( cusparseSetMatType( descrA, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatType( descrB, CUSPARSE_MATRIX_TYPE_GENERAL ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrA, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrB, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE(
cusparseScsr2csc( handle, A.num_rows, A.num_cols, A.nnz,
A.dval, A.drow, A.dcol, B->dval, B->dcol, B->drow,
CUSPARSE_ACTION_NUMERIC,
CUSPARSE_INDEX_BASE_ZERO) );
CHECK( magma_smconjugate( B, queue ));
} else if ( A.memory_location == Magma_CPU ){
CHECK( magma_smtransfer( A, &A_d, A.memory_location, Magma_DEV, queue ));
CHECK( magma_smtransposeconjugate( A_d, &B_d, queue ));
CHECK( magma_smtransfer( B_d, B, Magma_DEV, A.memory_location, queue ));
} else {
CHECK( magma_smconvert( A, &ACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_smtransposeconjugate( ACSR, &BCSR, queue ));
CHECK( magma_smconvert( BCSR, B, Magma_CSR, A.storage_type, queue ));
}
cleanup:
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroy( handle );
magma_smfree( &A_d, queue );
magma_smfree( &B_d, queue );
magma_smfree( &ACSR, queue );
magma_smfree( &BCSR, queue );
if( info != 0 ){
magma_smfree( B, queue );
}
return info;
}
extern "C" magma_int_t
magma_scsrsplit(
magma_int_t offset,
magma_int_t bsize,
magma_s_matrix A,
magma_s_matrix *D,
magma_s_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<offset; i+=offset ){
for( k=i; k<min(A.num_rows,i+offset); 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+offset ){
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", int(i), int(i));
info = -1;
goto cleanup;
}
}
}
magma_int_t ii = i;
for( i=ii; 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", int(i), int(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_smalloc_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_smalloc_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<offset; i+=offset) {
for( k=i; k<min(A.num_rows,i+offset); 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+offset ) {
// 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++;
}
}
}
}
ii = i;
for( i=ii; 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_s_matrix Ah={Magma_CSR}, ACSR={Magma_CSR}, DCSR={Magma_CSR}, RCSR={Magma_CSR}, Dh={Magma_CSR}, Rh={Magma_CSR};
CHECK( magma_smtransfer( A, &Ah, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smconvert( Ah, &ACSR, A.storage_type, Magma_CSR, queue ));
CHECK( magma_scsrsplit( offset, bsize, ACSR, &DCSR, &RCSR, queue ));
CHECK( magma_smconvert( DCSR, &Dh, Magma_CSR, A.storage_type, queue ));
CHECK( magma_smconvert( RCSR, &Rh, Magma_CSR, A.storage_type, queue ));
CHECK( magma_smtransfer( Dh, D, Magma_CPU, A.memory_location, queue ));
CHECK( magma_smtransfer( Rh, R, Magma_CPU, A.memory_location, queue ));
magma_smfree( &Ah, queue );
magma_smfree( &ACSR, queue );
magma_smfree( &Dh, queue );
magma_smfree( &DCSR, queue );
magma_smfree( &Rh, queue );
magma_smfree( &RCSR, queue );
}
cleanup:
if( info != 0 ){
magma_smfree( D, queue );
magma_smfree( R, queue );
}
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_MAKE(0.0, 0.0);
magma_s_matrix A={Magma_CSR}, B_d={Magma_CSR};
magma_s_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
magma_int_t n = A.num_rows;
CHECK( magma_smtransfer( A, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_svinit( &b, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_svinit( &x, Magma_DEV, A.num_cols, 1, one, queue ));
CHECK( magma_sprint_vector( b, 90, 10, queue ));
CHECK( magma_sprint_matrix( A, queue ));
printf("\n\n\n");
CHECK( magma_sprint_matrix( B_d, queue ));
float res;
res = magma_snrm2(n, b.dval, 1, queue );
printf("norm0: %f\n", res);
CHECK( magma_s_spmv( one, B_d, x, zero, b, queue )); // b = A x
CHECK( magma_sprint_vector( b, 0, 100, queue ));
CHECK( magma_sprint_vector( b, b.num_rows-10, 10, queue ));
res = magma_snrm2( n, b.dval, 1, queue );
printf("norm: %f\n", res);
CHECK( magma_sresidual( B_d, x, b, &res, queue));
printf("res: %f\n", res);
magma_smfree(&B_d, queue );
magma_smfree(&A, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
i++;
}
cleanup:
magma_smfree(&A, queue );
magma_smfree(&B_d, queue );
magma_smfree(&x, queue );
magma_smfree(&b, queue );
magma_queue_destroy( queue );
magma_finalize();
return info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing csr matrix add
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( &queue );
real_Double_t res;
magma_s_matrix A={Magma_CSR}, B={Magma_CSR}, B2={Magma_CSR},
A_d={Magma_CSR}, B_d={Magma_CSR}, C_d={Magma_CSR};
float one = MAGMA_S_MAKE(1.0, 0.0);
float mone = MAGMA_S_MAKE(-1.0, 0.0);
magma_int_t i=1;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
i++;
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_sm_5stencil( laplace_size, &B, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &B, argv[i], queue ));
}
printf("%% matrix info: %d-by-%d with %d nonzeros\n",
int(B.num_rows), int(B.num_cols), int(B.nnz) );
CHECK( magma_smtransfer( A, &A_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
CHECK( magma_scuspaxpy( &one, A_d, &one, B_d, &C_d, queue ));
magma_smfree(&B_d, queue );
CHECK( magma_scuspaxpy( &mone, A_d, &one, C_d, &B_d, queue ));
CHECK( magma_smtransfer( B_d, &B2, Magma_DEV, Magma_CPU, queue ));
magma_smfree(&A_d, queue );
magma_smfree(&B_d, queue );
magma_smfree(&C_d, queue );
// check difference
CHECK( magma_smdiff( B, B2, &res, queue ));
printf("%% ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("%% tester matrix add: ok\n");
else
printf("%% tester matrix add: failed\n");
magma_smfree(&A, queue );
magma_smfree(&B, queue );
magma_smfree(&B2, queue );
cleanup:
magma_smfree(&A_d, queue );
magma_smfree(&B_d, queue );
magma_smfree(&C_d, queue );
magma_smfree(&A, queue );
magma_smfree(&B, queue );
magma_smfree(&B2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return info;
}
extern "C" magma_int_t
magma_sjacobisetup_diagscal(
magma_s_matrix A, magma_s_matrix *d,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i;
magma_s_matrix A_h1={Magma_CSR}, B={Magma_CSR};
magma_s_matrix diag={Magma_CSR};
CHECK( magma_svinit( &diag, Magma_CPU, A.num_rows, 1, MAGMA_S_ZERO, queue ));
if ( A.storage_type != Magma_CSR || A.memory_location != Magma_CPU ) {
CHECK( magma_smtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue ));
for( magma_int_t rowindex=0; rowindex<B.num_rows; rowindex++ ) {
magma_int_t start = (B.drow[rowindex]);
magma_int_t end = (B.drow[rowindex+1]);
for( i=start; i<end; i++ ) {
if ( B.dcol[i]==rowindex ) {
diag.val[rowindex] = 1.0/B.val[i];
break;
}
}
if ( diag.val[rowindex] == MAGMA_S_ZERO ){
printf(" error: zero diagonal element in row %d!\n",
int(rowindex));
if ( A.storage_type != Magma_CSR) {
magma_smfree( &A_h1, queue );
}
magma_smfree( &B, queue );
magma_smfree( &diag, queue );
info = MAGMA_ERR_BADPRECOND;
goto cleanup;
}
}
}
else{
for( magma_int_t rowindex=0; rowindex<A.num_rows; rowindex++ ) {
magma_int_t start = (A.drow[rowindex]);
magma_int_t end = (A.drow[rowindex+1]);
for( i=start; i<end; i++ ) {
if ( A.dcol[i]==rowindex ) {
diag.val[rowindex] = 1.0/A.val[i];
break;
}
}
if ( diag.val[rowindex] == MAGMA_S_ZERO ){
printf(" error: zero diagonal element in row %d!\n",
int(rowindex));
if ( A.storage_type != Magma_CSR) {
magma_smfree( &A_h1, queue );
}
magma_smfree( &B, queue );
magma_smfree( &diag, queue );
info = MAGMA_ERR_BADPRECOND;
goto cleanup;
}
}
}
CHECK( magma_smtransfer( diag, d, Magma_CPU, Magma_DEV, queue ));
cleanup:
magma_smfree( &A_h1, queue );
magma_smfree( &B, queue );
magma_smfree( &diag, queue );
return info;
}
magma_int_t
magma_scustomicsetup(
magma_s_matrix A,
magma_s_matrix b,
magma_s_preconditioner *precond,
magma_queue_t queue )
{
magma_int_t info = 0;
cusparseHandle_t cusparseHandle=NULL;
cusparseMatDescr_t descrL=NULL;
cusparseMatDescr_t descrU=NULL;
magma_s_matrix hA={Magma_CSR};
char preconditionermatrix[255];
snprintf( preconditionermatrix, sizeof(preconditionermatrix),
"precondL.mtx" );
CHECK( magma_s_csr_mtx( &hA, preconditionermatrix , queue) );
// for CUSPARSE
CHECK( magma_smtransfer( hA, &precond->M, Magma_CPU, Magma_DEV , queue ));
// copy the matrix to precond->L and (transposed) to precond->U
CHECK( magma_smtransfer(precond->M, &(precond->L), Magma_DEV, Magma_DEV, queue ));
CHECK( magma_smtranspose( precond->L, &(precond->U), queue ));
// extract the diagonal of L into precond->d
CHECK( magma_sjacobisetup_diagscal( precond->L, &precond->d, queue ));
CHECK( magma_svinit( &precond->work1, Magma_DEV, hA.num_rows, 1, MAGMA_S_ZERO, queue ));
// extract the diagonal of U into precond->d2
CHECK( magma_sjacobisetup_diagscal( precond->U, &precond->d2, queue ));
CHECK( magma_svinit( &precond->work2, Magma_DEV, hA.num_rows, 1, MAGMA_S_ZERO, queue ));
// CUSPARSE context //
CHECK_CUSPARSE( cusparseCreate( &cusparseHandle ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrL ));
CHECK_CUSPARSE( cusparseSetMatType( descrL, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrL, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrL, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrL, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoL ));
CHECK_CUSPARSE( cusparseScsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrL,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoL ));
CHECK_CUSPARSE( cusparseCreateMatDescr( &descrU ));
CHECK_CUSPARSE( cusparseSetMatType( descrU, CUSPARSE_MATRIX_TYPE_TRIANGULAR ));
CHECK_CUSPARSE( cusparseSetMatDiagType( descrU, CUSPARSE_DIAG_TYPE_NON_UNIT ));
CHECK_CUSPARSE( cusparseSetMatIndexBase( descrU, CUSPARSE_INDEX_BASE_ZERO ));
CHECK_CUSPARSE( cusparseSetMatFillMode( descrU, CUSPARSE_FILL_MODE_LOWER ));
CHECK_CUSPARSE( cusparseCreateSolveAnalysisInfo( &precond->cuinfoU ));
CHECK_CUSPARSE( cusparseScsrsv_analysis( cusparseHandle,
CUSPARSE_OPERATION_TRANSPOSE, precond->M.num_rows,
precond->M.nnz, descrU,
precond->M.val, precond->M.row, precond->M.col, precond->cuinfoU ));
cleanup:
cusparseDestroy( cusparseHandle );
cusparseDestroyMatDescr( descrL );
cusparseDestroyMatDescr( descrU );
cusparseHandle=NULL;
descrL=NULL;
descrU=NULL;
magma_smfree( &hA, queue );
return info;
}
extern "C" magma_int_t
magma_sidr_strms(
magma_s_matrix A, magma_s_matrix b, magma_s_matrix *x,
magma_s_solver_par *solver_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_IDRMERGE;
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 float c_zero = MAGMA_S_ZERO;
const float c_one = MAGMA_S_ONE;
const float c_n_one = MAGMA_S_NEG_ONE;
// internal user options
const magma_int_t smoothing = 1; // 0 = disable, 1 = enable
const float 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;
magma_int_t ldd;
magma_int_t q;
float residual;
float nrm;
float nrmb;
float nrmr;
float nrmt;
float rho;
float om;
float gamma;
// matrices and vectors
magma_s_matrix dxs = {Magma_CSR};
magma_s_matrix dr = {Magma_CSR}, drs = {Magma_CSR};
magma_s_matrix dP = {Magma_CSR}, dP1 = {Magma_CSR};
magma_s_matrix dG = {Magma_CSR}, dGcol = {Magma_CSR};
magma_s_matrix dU = {Magma_CSR};
magma_s_matrix dM = {Magma_CSR};
magma_s_matrix df = {Magma_CSR};
magma_s_matrix dt = {Magma_CSR}, dtt = {Magma_CSR};
magma_s_matrix dc = {Magma_CSR};
magma_s_matrix dv = {Magma_CSR};
magma_s_matrix dskp = {Magma_CSR};
magma_s_matrix dalpha = {Magma_CSR};
magma_s_matrix dbeta = {Magma_CSR};
float *hMdiag = NULL;
float *hskp = NULL;
float *halpha = NULL;
float *hbeta = NULL;
float *d1 = NULL, *d2 = NULL;
// queue variables
const magma_int_t nqueues = 3; // number of queues
magma_queue_t queues[nqueues];
// chronometry
real_Double_t tempo1, tempo2;
// create additional queues
queues[0] = queue;
for ( q = 1; q < nqueues; q++ ) {
magma_queue_create( queue->device(), &(queues[q]) );
}
// 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_snrm2( b.num_rows, b.dval, 1, queue );
if ( nrmb == 0.0 ) {
magma_sscal( x->num_rows, MAGMA_S_ZERO, x->dval, 1, queue );
info = MAGMA_SUCCESS;
goto cleanup;
}
// t = 0
// make t twice as large to contain both, dt and dr
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_svinit( &dt, Magma_DEV, ldd, 2, c_zero, queue ));
dt.num_rows = b.num_rows;
dt.num_cols = 1;
dt.nnz = dt.num_rows;
// redirect the dr.dval to the second part of dt
CHECK( magma_svinit( &dr, Magma_DEV, b.num_rows, 1, c_zero, queue ));
magma_free( dr.dval );
dr.dval = dt.dval + ldd;
// r = b - A x
CHECK( magma_sresidualvec( 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_svinit( &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_slarnv( &distr, iseed, &dof, dP.val );
// transfer P to device
CHECK( magma_smtransfer( dP, &dP1, Magma_CPU, Magma_DEV, queue ));
magma_smfree( &dP, queue );
// P = ortho(P1)
if ( dP1.num_cols > 1 ) {
// P = magma_sqr(P1), QR factorization
CHECK( magma_sqr( dP1.num_rows, dP1.num_cols, dP1, dP1.ld, &dP, NULL, queue ));
} else {
// P = P1 / |P1|
nrm = magma_snrm2( dof, dP1.dval, 1, queue );
nrm = 1.0 / nrm;
magma_sscal( dof, nrm, dP1.dval, 1, queue );
CHECK( magma_smtransfer( dP1, &dP, Magma_DEV, Magma_DEV, queue ));
}
magma_smfree( &dP1, queue );
//---------------------------------------
// allocate memory for the scalar products
CHECK( magma_smalloc_pinned( &hskp, 5 ));
CHECK( magma_svinit( &dskp, Magma_DEV, 4, 1, c_zero, queue ));
CHECK( magma_smalloc_pinned( &halpha, s ));
CHECK( magma_svinit( &dalpha, Magma_DEV, s, 1, c_zero, queue ));
CHECK( magma_smalloc_pinned( &hbeta, s ));
CHECK( magma_svinit( &dbeta, Magma_DEV, s, 1, c_zero, queue ));
// workspace for merged dot product
CHECK( magma_smalloc( &d1, max(2, s) * b.num_rows ));
CHECK( magma_smalloc( &d2, max(2, s) * b.num_rows ));
// smoothing enabled
if ( smoothing > 0 ) {
// set smoothing solution vector
CHECK( magma_smtransfer( *x, &dxs, Magma_DEV, Magma_DEV, queue ));
// tt = 0
// make tt twice as large to contain both, dtt and drs
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_svinit( &dtt, Magma_DEV, ldd, 2, c_zero, queue ));
dtt.num_rows = dr.num_rows;
dtt.num_cols = 1;
dtt.nnz = dtt.num_rows;
// redirect the drs.dval to the second part of dtt
CHECK( magma_svinit( &drs, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
magma_free( drs.dval );
drs.dval = dtt.dval + ldd;
// set smoothing residual vector
magma_scopyvector( dr.num_rows, dr.dval, 1, drs.dval, 1, queue );
}
// G(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_rows, 32 );
CHECK( magma_svinit( &dG, Magma_DEV, ldd, s, c_zero, queue ));
dG.num_rows = A.num_rows;
} else {
CHECK( magma_svinit( &dG, Magma_DEV, A.num_rows, s, c_zero, queue ));
}
// dGcol represents a single column of dG, array pointer is set inside loop
CHECK( magma_svinit( &dGcol, Magma_DEV, dG.num_rows, 1, c_zero, queue ));
magma_free( dGcol.dval );
// U(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_cols, 32 );
CHECK( magma_svinit( &dU, Magma_DEV, ldd, s, c_zero, queue ));
dU.num_rows = A.num_cols;
} else {
CHECK( magma_svinit( &dU, Magma_DEV, A.num_cols, s, c_zero, queue ));
}
// M(s,s) = I
CHECK( magma_svinit( &dM, Magma_DEV, s, s, c_zero, queue ));
CHECK( magma_smalloc_pinned( &hMdiag, s ));
magmablas_slaset( MagmaFull, dM.num_rows, dM.num_cols, c_zero, c_one, dM.dval, dM.ld, queue );
// f = 0
CHECK( magma_svinit( &df, Magma_DEV, dP.num_cols, 1, c_zero, queue ));
// c = 0
CHECK( magma_svinit( &dc, Magma_DEV, dM.num_cols, 1, c_zero, queue ));
// v = r
CHECK( magma_smtransfer( dr, &dv, Magma_DEV, Magma_DEV, queue ));
//--------------START TIME---------------
// chronometry
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->timing[0] = 0.0;
}
cudaProfilerStart();
om = MAGMA_S_ONE;
gamma = MAGMA_S_ZERO;
innerflag = 0;
// new RHS for small systems
// f = P' r
// Q1
magma_sgemvmdot_shfl( dP.num_rows, dP.num_cols, dP.dval, dr.dval, d1, d2, df.dval, queues[1] );
// skp[4] = f(k)
// Q1
magma_sgetvector_async( 1, df.dval, 1, &hskp[4], 1, queues[1] );
// c(k:s) = f(k:s)
// Q1
magma_scopyvector_async( s, df.dval, 1, dc.dval, 1, queues[1] );
// c(k:s) = M(k:s,k:s) \ f(k:s)
// Q1
magma_strsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, s, dM.dval, dM.ld, dc.dval, 1, queues[1] );
// start iteration
do
{
solver_par->numiter++;
// shadow space loop
for ( k = 0; k < s; ++k ) {
sk = s - k;
dGcol.dval = dG.dval + k * dG.ld;
// v = r - G(:,k:s) c(k:s)
// Q1
magmablas_sgemv( MagmaNoTrans, dG.num_rows, sk, c_n_one, dGcol.dval, dG.ld, &dc.dval[k], 1, c_one, dv.dval, 1, queues[1] );
// U(:,k) = om * v + U(:,k:s) c(k:s)
// Q1
magmablas_sgemv( MagmaNoTrans, dU.num_rows, sk, c_one, &dU.dval[k*dU.ld], dU.ld, &dc.dval[k], 1, om, dv.dval, 1, queues[1] );
// G(:,k) = A U(:,k)
// Q1
CHECK( magma_s_spmv( c_one, A, dv, c_zero, dGcol, queues[1] ));
solver_par->spmv_count++;
// bi-orthogonalize the new basis vectors
for ( i = 0; i < k; ++i ) {
// alpha = P(:,i)' G(:,k)
// Q1
halpha[i] = magma_sdot( dP.num_rows, &dP.dval[i*dP.ld], 1, dGcol.dval, 1, queues[1] );
// implicit sync Q1 --> alpha = P(:,i)' G(:,k)
// alpha = alpha / M(i,i)
halpha[i] = halpha[i] / hMdiag[i];
// G(:,k) = G(:,k) - alpha * G(:,i)
// Q1
magma_saxpy( dG.num_rows, -halpha[i], &dG.dval[i*dG.ld], 1, dGcol.dval, 1, queues[1] );
}
// sync Q1 --> G(:,k) = G(:,k) - alpha * G(:,i), skp[4] = f(k)
magma_queue_sync( queues[1] );
// new column of M = P'G, first k-1 entries are zero
// M(k:s,k) = P(:,k:s)' G(:,k)
// Q2
magma_sgemvmdot_shfl( dP.num_rows, sk, &dP.dval[k*dP.ld], dGcol.dval, d1, d2, &dM.dval[k*dM.ld+k], queues[2] );
// non-first s iteration
if ( k > 0 ) {
// alpha = dalpha
// Q0
magma_ssetvector_async( k, halpha, 1, dalpha.dval, 1, queues[0] );
// U update outside of loop using GEMV
// U(:,k) = U(:,k) - U(:,1:k) * alpha(1:k)
// Q0
magmablas_sgemv( MagmaNoTrans, dU.num_rows, k, c_n_one, dU.dval, dU.ld, dalpha.dval, 1, c_one, dv.dval, 1, queues[0] );
}
// Mdiag(k) = M(k,k)
// Q2
magma_sgetvector( 1, &dM.dval[k*dM.ld+k], 1, &hMdiag[k], 1, queues[2] );
// implicit sync Q2 --> Mdiag(k) = M(k,k)
// U(:,k) = v
// Q0
magma_scopyvector_async( dU.num_rows, dv.dval, 1, &dU.dval[k*dU.ld], 1, queues[0] );
// check M(k,k) == 0
if ( MAGMA_S_EQUAL(hMdiag[k], MAGMA_S_ZERO) ) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// beta = f(k) / M(k,k)
hbeta[k] = hskp[4] / hMdiag[k];
// check for nan
if ( magma_s_isnan( hbeta[k] ) || magma_s_isinf( hbeta[k] )) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// r = r - beta * G(:,k)
// Q2
magma_saxpy( dr.num_rows, -hbeta[k], dGcol.dval, 1, dr.dval, 1, queues[2] );
// non-last s iteration
if ( (k + 1) < s ) {
// f(k+1:s) = f(k+1:s) - beta * M(k+1:s,k)
// Q1
magma_saxpy( sk-1, -hbeta[k], &dM.dval[k*dM.ld+(k+1)], 1, &df.dval[k+1], 1, queues[1] );
// c(k+1:s) = f(k+1:s)
// Q1
magma_scopyvector_async( sk-1, &df.dval[k+1], 1, &dc.dval[k+1], 1, queues[1] );
// c(k+1:s) = M(k+1:s,k+1:s) \ f(k+1:s)
// Q1
magma_strsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, sk-1, &dM.dval[(k+1)*dM.ld+(k+1)], dM.ld, &dc.dval[k+1], 1, queues[1] );
// skp[4] = f(k+1)
// Q1
magma_sgetvector_async( 1, &df.dval[k+1], 1, &hskp[4], 1, queues[1] );
}
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
// Q2
nrmr = magma_snrm2( dr.num_rows, dr.dval, 1, queues[2] );
// implicit sync Q2 --> |r|
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
// Q2
magma_sidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queues[2] );
// x = x + beta * U(:,k)
// Q0
magma_saxpy( x->num_rows, hbeta[k], &dU.dval[k*dU.ld], 1, x->dval, 1, queues[0] );
// t't
// t'rs
// Q2
CHECK( magma_sgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queues[2] ));
// skp[2-3] = dskp[2-3]
// Q2
magma_sgetvector( 2, &dskp.dval[2], 1, &hskp[2], 1, queues[2] );
// implicit sync Q2 --> skp = dskp
// gamma = (t' * rs) / (t' * t)
gamma = hskp[3] / hskp[2];
// rs = rs - gamma * t
// Q1
magma_saxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queues[1] );
// xs = xs - gamma * (xs - x)
// Q0
magma_sidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queues[0] );
// |rs|
// Q1
nrmr = magma_snrm2( drs.num_rows, drs.dval, 1, queues[1] );
// implicit sync Q0 --> |r|
//---------------------------------------
}
// v = r
// Q1
magma_scopyvector_async( dr.num_rows, dr.dval, 1, dv.dval, 1, queues[1] );
// last s iteration
if ( (k + 1) == s ) {
// t = A r
// Q2
CHECK( magma_s_spmv( c_one, A, dr, c_zero, dt, queues[2] ));
solver_par->spmv_count++;
// t't
// t'r
// Q2
CHECK( magma_sgemvmdot_shfl( dt.ld, 2, dt.dval, dt.dval, d1, d2, dskp.dval, queues[2] ));
}
// 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 or iteration limit
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;
}
}
// smoothing disabled
if ( smoothing <= 0 && innerflag != 1 ) {
// dbeta(1:s) = beta(1:s)
// Q0
magma_ssetvector_async( s, hbeta, 1, dbeta.dval, 1, queues[0] );
// x = x + U(:,1:s) * beta(1:s)
// Q0
magmablas_sgemv( MagmaNoTrans, dU.num_rows, s, c_one, dU.dval, dU.ld, dbeta.dval, 1, c_one, x->dval, 1, queues[0] );
}
// check convergence or iteration limit or invalid result of inner loop
if ( innerflag > 0 ) {
break;
}
// computation of a new omega
//---------------------------------------
// skp[0-2] = dskp[0-2]
// Q2
magma_sgetvector( 2, dskp.dval, 1, hskp, 1, queues[2] );
// implicit sync Q2 --> skp = dskp
// |t|
nrmt = magma_ssqrt( MAGMA_S_REAL(hskp[0]) );
// rho = abs((t' * r) / (|t| * |r|))
rho = MAGMA_D_ABS( MAGMA_S_REAL(hskp[1]) / (nrmt * nrmr) );
// om = (t' * r) / (|t| * |t|)
om = hskp[1] / hskp[0];
if ( rho < angle ) {
om = (om * angle) / rho;
}
//---------------------------------------
if ( MAGMA_S_EQUAL(om, MAGMA_S_ZERO) ) {
info = MAGMA_DIVERGENCE;
break;
}
// sync Q1 --> v = r
magma_queue_sync( queues[1] );
// r = r - om * t
// Q2
magma_saxpy( dr.num_rows, -om, dt.dval, 1, dr.dval, 1, queues[2] );
// x = x + om * v
// Q0
magma_saxpy( x->num_rows, om, dv.dval, 1, x->dval, 1, queues[0] );
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
// Q2
nrmr = magma_snrm2( dr.num_rows, dr.dval, 1, queues[2] );
// implicit sync Q2 --> |r|
// v = r
// Q0
magma_scopyvector_async( dr.num_rows, dr.dval, 1, dv.dval, 1, queues[0] );
// new RHS for small systems
// f = P' r
// Q1
magma_sgemvmdot_shfl( dP.num_rows, dP.num_cols, dP.dval, dr.dval, d1, d2, df.dval, queues[1] );
// skp[4] = f(k)
// Q1
magma_sgetvector_async( 1, df.dval, 1, &hskp[4], 1, queues[1] );
// c(k:s) = f(k:s)
// Q1
magma_scopyvector_async( s, df.dval, 1, dc.dval, 1, queues[1] );
// c(k:s) = M(k:s,k:s) \ f(k:s)
// Q1
magma_strsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, s, dM.dval, dM.ld, dc.dval, 1, queues[1] );
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
// Q2
magma_sidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queues[2] );
// t't
// t'rs
// Q2
CHECK( magma_sgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queues[2] ));
// skp[2-3] = dskp[2-3]
// Q2
magma_sgetvector( 2, &dskp.dval[2], 1, &hskp[2], 1, queues[2] );
// implicit sync Q2 --> skp = dskp
// gamma = (t' * rs) / (t' * t)
gamma = hskp[3] / hskp[2];
// rs = rs - gamma * (rs - r)
// Q2
magma_saxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queues[2] );
// xs = xs - gamma * (xs - x)
// Q0
magma_sidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queues[0] );
// v = r
// Q0
magma_scopyvector_async( dr.num_rows, dr.dval, 1, dv.dval, 1, queues[0] );
// new RHS for small systems
// f = P' r
// Q1
magma_sgemvmdot_shfl( dP.num_rows, dP.num_cols, dP.dval, dr.dval, d1, d2, df.dval, queues[1] );
// skp[4] = f(k)
// Q1
magma_sgetvector_async( 1, df.dval, 1, &hskp[4], 1, queues[1] );
// c(k:s) = f(k:s)
// Q1
magma_scopyvector_async( s, df.dval, 1, dc.dval, 1, queues[1] );
// |rs|
// Q2
nrmr = magma_snrm2( drs.num_rows, drs.dval, 1, queues[2] );
// implicit sync Q2 --> |r|
// c(k:s) = M(k:s,k:s) \ f(k:s)
// Q1
magma_strsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, s, dM.dval, dM.ld, dc.dval, 1, queues[1] );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
magma_queue_sync( 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 or iteration limit
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
break;
}
// sync Q0 --> v = r
magma_queue_sync( queues[0] );
}
while ( solver_par->numiter + 1 <= solver_par->maxiter );
// sync all queues
for ( q = 0; q < nqueues; q++ ) {
magma_queue_sync( queues[q] );
}
// smoothing enabled
if ( smoothing > 0 ) {
// x = xs
magma_scopyvector_async( x->num_rows, dxs.dval, 1, x->dval, 1, queue );
// r = rs
magma_scopyvector_async( dr.num_rows, drs.dval, 1, dr.dval, 1, queue );
}
cudaProfilerStop();
// get last iteration timing
tempo2 = magma_sync_wtime( queue );
magma_queue_sync( queue );
solver_par->runtime = (real_Double_t)tempo2 - tempo1;
//--------------STOP TIME----------------
// get final stats
solver_par->iter_res = nrmr;
CHECK( magma_sresidualvec( 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
// sync all queues, destory additional queues
magma_queue_sync( queues[0] );
for ( q = 1; q < nqueues; q++ ) {
magma_queue_sync( queues[q] );
magma_queue_destroy( queues[q] );
}
// smoothing enabled
if ( smoothing > 0 ) {
drs.dval = NULL; // needed because its pointer is redirected to dtt
magma_smfree( &dxs, queue );
magma_smfree( &drs, queue );
magma_smfree( &dtt, queue );
}
dr.dval = NULL; // needed because its pointer is redirected to dt
dGcol.dval = NULL; // needed because its pointer is redirected to dG
magma_smfree( &dr, queue );
magma_smfree( &dP, queue );
magma_smfree( &dP1, queue );
magma_smfree( &dG, queue );
magma_smfree( &dGcol, queue );
magma_smfree( &dU, queue );
magma_smfree( &dM, queue );
magma_smfree( &df, queue );
magma_smfree( &dt, queue );
magma_smfree( &dc, queue );
magma_smfree( &dv, queue );
magma_smfree( &dskp, queue );
magma_smfree( &dalpha, queue );
magma_smfree( &dbeta, queue );
magma_free_pinned( hMdiag );
magma_free_pinned( hskp );
magma_free_pinned( halpha );
magma_free_pinned( hbeta );
magma_free( d1 );
magma_free( d2 );
solver_par->info = info;
return info;
/* magma_sidr_strms */
}
extern "C" magma_int_t
magma_slsqr(
magma_s_matrix A, magma_s_matrix b, magma_s_matrix *x,
magma_s_solver_par *solver_par,
magma_s_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_LSQR;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
magma_int_t m = A.num_rows * b.num_cols;
magma_int_t n = A.num_cols * b.num_cols;
// local variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_ONE;
// solver variables
float s, nom0, r0, res=0, nomb, phibar, beta, alpha, c, rho, rhot, phi, thet, normr, normar, norma, sumnormd2, normd;
// need to transpose the matrix
magma_s_matrix AT={Magma_CSR}, Ah1={Magma_CSR}, Ah2={Magma_CSR};
// GPU workspace
magma_s_matrix r={Magma_CSR},
v={Magma_CSR}, z={Magma_CSR}, zt={Magma_CSR},
d={Magma_CSR}, vt={Magma_CSR}, q={Magma_CSR},
w={Magma_CSR}, u={Magma_CSR};
CHECK( magma_svinit( &r, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &v, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &z, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &d, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &vt,Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &q, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &w, Magma_DEV, A.num_cols, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &u, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_svinit( &zt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// transpose the matrix
magma_smtransfer( A, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_smconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_smfree(&Ah1, queue );
magma_smtransposeconjugate( Ah2, &Ah1, queue );
magma_smfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_smconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_smfree(&Ah1, queue );
magma_smtransfer( Ah2, &AT, Magma_CPU, Magma_DEV, queue );
magma_smfree(&Ah2, queue );
// solver setup
CHECK( magma_sresidualvec( A, b, *x, &r, &nom0, queue));
solver_par->init_res = nom0;
nomb = magma_snrm2( m, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
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)nom0;
solver_par->timing[0] = 0.0;
}
if ( nom0 < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
magma_scopy( m, b.dval, 1, u.dval, 1, queue );
beta = magma_snrm2( m, u.dval, 1, queue );
magma_sscal( m, MAGMA_S_MAKE(1./beta, 0.0 ), u.dval, 1, queue );
normr = beta;
c = 1.0;
s = 0.0;
phibar = beta;
CHECK( magma_s_spmv( c_one, AT, u, c_zero, v, queue ));
if( precond_par->solver == Magma_NONE ){
;
} else {
CHECK( magma_s_applyprecond_right( MagmaTrans, A, v, &zt, precond_par, queue ));
CHECK( magma_s_applyprecond_left( MagmaTrans, A, zt, &v, precond_par, queue ));
}
alpha = magma_snrm2( n, v.dval, 1, queue );
magma_sscal( n, MAGMA_S_MAKE(1./alpha, 0.0 ), v.dval, 1, queue );
normar = alpha * beta;
norma = 0;
sumnormd2 = 0;
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
// start iteration
do
{
solver_par->numiter++;
if( precond_par->solver == Magma_NONE || A.num_rows != A.num_cols ) {
magma_scopy( n, v.dval, 1 , z.dval, 1, queue );
} else {
CHECK( magma_s_applyprecond_left( MagmaNoTrans, A, v, &zt, precond_par, queue ));
CHECK( magma_s_applyprecond_right( MagmaNoTrans, A, zt, &z, precond_par, queue ));
}
//CHECK( magma_s_spmv( c_one, A, z, MAGMA_S_MAKE(-alpha,0.0), u, queue ));
CHECK( magma_s_spmv( c_one, A, z, c_zero, zt, queue ));
magma_sscal( m, MAGMA_S_MAKE(-alpha, 0.0 ), u.dval, 1, queue );
magma_saxpy( m, c_one, zt.dval, 1, u.dval, 1, queue );
solver_par->spmv_count++;
beta = magma_snrm2( m, u.dval, 1, queue );
magma_sscal( m, MAGMA_S_MAKE(1./beta, 0.0 ), u.dval, 1, queue );
// norma = norm([norma alpha beta]);
norma = sqrt(norma*norma + alpha*alpha + beta*beta );
//lsvec( solver_par->numiter-1 ) = normar / norma;
thet = -s * alpha;
rhot = c * alpha;
rho = sqrt( rhot * rhot + beta * beta );
c = rhot / rho;
s = - beta / rho;
phi = c * phibar;
phibar = s * phibar;
// d = (z - thet * d) / rho;
magma_sscal( n, MAGMA_S_MAKE(-thet, 0.0 ), d.dval, 1, queue );
magma_saxpy( n, c_one, z.dval, 1, d.dval, 1, queue );
magma_sscal( n, MAGMA_S_MAKE(1./rho, 0.0 ), d.dval, 1, queue );
normd = magma_snrm2( n, d.dval, 1, queue );
sumnormd2 = sumnormd2 + normd*normd;
// convergence check
res = normr;
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
// check for convergence in A*x=b
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
info = MAGMA_SUCCESS;
break;
}
// check for convergence in min{|b-A*x|}
if ( A.num_rows != A.num_cols &&
( normar/(norma*normr) <= solver_par->rtol || normar <= solver_par->atol ) ){
printf("%% warning: quit from minimization convergence check.\n");
info = MAGMA_SUCCESS;
break;
}
magma_saxpy( n, MAGMA_S_MAKE( phi, 0.0 ), d.dval, 1, x->dval, 1, queue );
normr = fabs(s) * normr;
CHECK( magma_s_spmv( c_one, AT, u, c_zero, vt, queue ));
solver_par->spmv_count++;
if( precond_par->solver == Magma_NONE ){
;
} else {
CHECK( magma_s_applyprecond_right( MagmaTrans, A, vt, &zt, precond_par, queue ));
CHECK( magma_s_applyprecond_left( MagmaTrans, A, zt, &vt, precond_par, queue ));
}
magma_sscal( n, MAGMA_S_MAKE(-beta, 0.0 ), v.dval, 1, queue );
magma_saxpy( n, c_one, vt.dval, 1, v.dval, 1, queue );
alpha = magma_snrm2( n, v.dval, 1, queue );
magma_sscal( n, MAGMA_S_MAKE(1./alpha, 0.0 ), v.dval, 1, queue );
normar = alpha * fabs(s*phi);
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
CHECK( magma_sresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter && info == MAGMA_SUCCESS ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_smfree(&r, queue );
magma_smfree(&v, queue );
magma_smfree(&z, queue );
magma_smfree(&zt, queue );
magma_smfree(&d, queue );
magma_smfree(&vt, queue );
magma_smfree(&q, queue );
magma_smfree(&u, queue );
magma_smfree(&w, queue );
magma_smfree(&AT, queue );
magma_smfree(&Ah1, queue );
magma_smfree(&Ah2, queue );
solver_par->info = info;
return info;
} /* magma_sqmr */
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_sopts zopts;
magma_queue_t queue=NULL;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
real_Double_t res;
magma_s_matrix A={Magma_CSR}, AT={Magma_CSR}, A2={Magma_CSR},
B={Magma_CSR}, B_d={Magma_CSR};
int i=1;
real_Double_t start, end;
CHECK( magma_sparse_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_sm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_s_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) A.num_rows,(int) A.num_cols,(int) A.nnz );
// scale matrix
CHECK( magma_smscale( &A, zopts.scaling, queue ));
// remove nonzeros in matrix
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_smcsrcompressor( &A, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA CPU: %.2e seconds.\n", (end-start)/10 );
// transpose
CHECK( magma_smtranspose( A, &AT, queue ));
// convert, copy back and forth to check everything works
CHECK( magma_smconvert( AT, &B, Magma_CSR, Magma_CSR, queue ));
magma_smfree(&AT, queue );
CHECK( magma_smtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue ));
magma_smfree(&B, queue );
start = magma_sync_wtime( queue );
for (int j=0; j<10; j++)
CHECK( magma_smcsrcompressor_gpu( &B_d, queue ));
end = magma_sync_wtime( queue );
printf( " > MAGMA GPU: %.2e seconds.\n", (end-start)/10 );
CHECK( magma_smtransfer( B_d, &B, Magma_DEV, Magma_CPU, queue ));
magma_smfree(&B_d, queue );
CHECK( magma_smconvert( B, &AT, Magma_CSR, Magma_CSR, queue ));
magma_smfree(&B, queue );
// transpose back
CHECK( magma_smtranspose( AT, &A2, queue ));
magma_smfree(&AT, queue );
CHECK( magma_smdiff( A, A2, &res, queue ));
printf("# ||A-B||_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester matrix compressor: ok\n");
else
printf("# tester matrix compressor: failed\n");
magma_smfree(&A, queue );
magma_smfree(&A2, queue );
i++;
}
cleanup:
magma_smfree(&AT, queue );
magma_smfree(&B, queue );
magma_smfree(&A, queue );
magma_smfree(&A2, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
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;
}
extern "C" magma_int_t
magma_sjacobisetup(
magma_s_matrix A, magma_s_matrix b,
magma_s_matrix *M, magma_s_matrix *c,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i;
magma_s_matrix A_h1={Magma_CSR}, A_h2={Magma_CSR}, B={Magma_CSR}, C={Magma_CSR};
magma_s_matrix diag={Magma_CSR}, c_t={Magma_CSR}, b_h={Magma_CSR};
CHECK( magma_svinit( &c_t, Magma_CPU, A.num_rows, b.num_cols, MAGMA_S_ZERO, queue ));
CHECK( magma_svinit( &diag, Magma_CPU, A.num_rows, b.num_cols, MAGMA_S_ZERO, queue ));
CHECK( magma_smtransfer( b, &b_h, A.memory_location, Magma_CPU, queue ));
if ( A.storage_type != Magma_CSR ) {
CHECK( magma_smtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue ));
CHECK( magma_smconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue ));
}
else {
CHECK( magma_smtransfer( A, &B, A.memory_location, Magma_CPU, queue ));
}
for( magma_int_t rowindex=0; rowindex<B.num_rows; rowindex++ ) {
magma_int_t start = (B.drow[rowindex]);
magma_int_t end = (B.drow[rowindex+1]);
for( i=start; i<end; i++ ) {
if ( B.dcol[i]==rowindex ) {
diag.val[rowindex] = B.val[i];
if ( MAGMA_S_REAL( diag.val[rowindex]) == 0 )
printf(" error: zero diagonal element in row %d!\n",
int(rowindex));
}
}
for( i=start; i<end; i++ ) {
B.val[i] = B.val[i] / diag.val[rowindex];
if ( B.dcol[i]==rowindex ) {
B.val[i] = MAGMA_S_MAKE( 0., 0. );
}
}
c_t.val[rowindex] = b_h.val[rowindex] / diag.val[rowindex];
}
CHECK( magma_s_csr_compressor(&B.val, &B.drow, &B.dcol,
&C.val, &C.drow, &C.dcol, &B.num_rows, queue ));
C.num_rows = B.num_rows;
C.num_cols = B.num_cols;
C.memory_location = B.memory_location;
C.nnz = C.drow[B.num_rows];
C.storage_type = B.storage_type;
C.memory_location = B.memory_location;
if ( A.storage_type != Magma_CSR) {
A_h2.alignment = A.alignment;
A_h2.blocksize = A.blocksize;
CHECK( magma_smconvert( C, &A_h2, Magma_CSR, A_h1.storage_type, queue ));
CHECK( magma_smtransfer( A_h2, M, Magma_CPU, A.memory_location, queue ));
}
else {
CHECK( magma_smtransfer( C, M, Magma_CPU, A.memory_location, queue ));
}
CHECK( magma_smtransfer( c_t, c, Magma_CPU, A.memory_location, queue ));
if ( A.storage_type != Magma_CSR) {
magma_smfree( &A_h1, queue );
magma_smfree( &A_h2, queue );
}
cleanup:
magma_smfree( &B, queue );
magma_smfree( &C, queue );
magma_smfree( &diag, queue );
magma_smfree( &c_t, queue );
magma_smfree( &b_h, queue );
return info;
}
