magma_int_t
magma_c_applyprecond( magma_c_sparse_matrix A, magma_c_vector b,
magma_c_vector *x, magma_c_preconditioner *precond )
{
if( precond->solver == Magma_JACOBI ){
magma_cjacobi_diagscal( A.num_rows, precond->d.val, b.val, x->val );
return MAGMA_SUCCESS;
}
else if( precond->solver == Magma_PASTIX ){
magma_capplypastix( b, x, precond );
return MAGMA_SUCCESS;
}
else if( precond->solver == Magma_ILU ){
magma_c_vector tmp;
magma_c_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_C_MAKE(1.0, 0.0) );
// magma_capplycuilu_l( b, &tmp, precond );
// magma_capplycuilu_r( tmp, x, precond );
magma_c_vfree( &tmp );
return MAGMA_SUCCESS;
}
else if( precond->solver == Magma_ICC ){
magma_c_vector tmp;
magma_c_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_C_MAKE(1.0, 0.0) );
// magma_ctrisv_l_nu( precond->L, b, &tmp );
// magma_ctrisv_r_nu( precond->L, tmp, x );
magma_c_vfree( &tmp );
return MAGMA_SUCCESS;
}
else{
printf( "error: preconditioner type not yet supported.\n" );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
extern "C" magma_int_t
magma_c_applyprecond(
magma_c_sparse_matrix A,
magma_c_vector b,
magma_c_vector *x,
magma_c_preconditioner *precond,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
if ( precond->solver == Magma_JACOBI ) {
magma_cjacobi_diagscal( A.num_rows, precond->d, b, x, queue );
}
else if ( precond->solver == Magma_PASTIX ) {
magma_capplypastix( b, x, precond, queue );
}
else if ( precond->solver == Magma_ILU ) {
magma_c_vector tmp;
magma_c_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_C_ZERO, queue );
magma_c_vfree( &tmp, queue );
}
else if ( precond->solver == Magma_ICC ) {
magma_c_vector tmp;
magma_c_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_C_ZERO, queue );
magma_c_vfree( &tmp, queue );
}
else if ( precond->solver == Magma_NONE ) {
magma_ccopy( b.num_rows, b.dval, 1, x->dval, 1 ); // x = b
}
else {
printf( "error: preconditioner type not yet supported.\n" );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_copts zopts;
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
int i=1;
magma_cparse_opts( argc, argv, &zopts, &i, queue );
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_sparse_matrix A, B, B_d;
magma_c_vector x, b;
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.precond_par.solver = Magma_NONE;
magma_csolverinfo_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] );
magma_cm_5stencil( laplace_size, &A, queue );
} else { // file-matrix test
magma_c_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;
magma_ceigensolverinfo_init( &zopts.solver_par, queue );
// scale matrix
magma_cmscale( &A, zopts.scaling, queue );
magma_c_mconvert( A, &B, Magma_CSR, zopts.output_format, queue );
magma_c_mtransfer( B, &B_d, Magma_CPU, Magma_DEV, queue );
// vectors and initial guess
magma_c_vinit( &b, Magma_DEV, A.num_cols, one, queue );
magma_c_vinit( &x, Magma_DEV, A.num_cols, one, queue );
magma_c_spmv( one, B_d, x, zero, b, queue ); // b = A x
magma_c_vfree(&x, queue );
magma_c_vinit( &x, Magma_DEV, A.num_cols, zero, queue );
magma_c_solver( B_d, b, &x, &zopts, queue );
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par, queue );
magma_c_mfree(&B_d, queue );
magma_c_mfree(&B, queue );
magma_c_mfree(&A, queue );
magma_c_vfree(&x, queue );
magma_c_vfree(&b, queue );
i++;
}
magma_csolverinfo_free( &zopts.solver_par, &zopts.precond_par, queue );
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
magma_int_t
magma_cpastixsetup( magma_c_sparse_matrix A, magma_c_vector b,
magma_c_preconditioner *precond ){
#if defined(HAVE_PASTIX)
#if defined(PRECISION_d)
pastix_data_t *pastix_data = NULL; /* Pointer to a storage structure needed by pastix */
pastix_int_t ncol; /* Size of the matrix */
pastix_int_t *colptr = NULL; /* Indexes of first element of each column in row and values */
pastix_int_t *rows = NULL; /* Row of each element of the matrix */
pastix_float_t *values = NULL; /* Value of each element of the matrix */
pastix_float_t *rhs = NULL; /* right hand side */
pastix_int_t *iparm = NULL; /* integer parameters for pastix */
float *dparm = NULL; /* floating parameters for pastix */
pastix_int_t *perm = NULL; /* Permutation tabular */
pastix_int_t *invp = NULL; /* Reverse permutation tabular */
pastix_int_t mat_type;
magma_c_sparse_matrix A_h1, B;
magma_c_vector diag, c_t, b_h;
magma_c_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_C_ZERO );
magma_c_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_C_ZERO );
magma_c_vtransfer( b, &b_h, A.memory_location, Magma_CPU);
if( A.storage_type != Magma_CSR ){
magma_c_mtransfer( A, &A_h1, A.memory_location, Magma_CPU);
magma_c_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR);
}
else{
magma_c_mtransfer( A, &B, A.memory_location, Magma_CPU);
}
rhs = (pastix_float_t*) b_h.val;
ncol = B.num_rows;
colptr = B.row;
rows = B.col;
values = (pastix_float_t*) B.val;
mat_type = API_SYM_NO;
iparm = (pastix_int_t*)malloc(IPARM_SIZE*sizeof(pastix_int_t));
dparm = (pastix_float_t*)malloc(DPARM_SIZE*sizeof(pastix_float_t));
/*******************************************/
/* Initialize parameters to default values */
/*******************************************/
iparm[IPARM_MODIFY_PARAMETER] = API_NO;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, rhs, 1, iparm, dparm);
iparm[IPARM_THREAD_NBR] = 16;
iparm[IPARM_SYM] = mat_type;
iparm[IPARM_FACTORIZATION] = API_FACT_LU;
iparm[IPARM_VERBOSE] = API_VERBOSE_YES;
iparm[IPARM_ORDERING] = API_ORDER_SCOTCH;
iparm[IPARM_INCOMPLETE] = API_NO;
iparm[IPARM_RHS_MAKING] = API_RHS_B;
//iparm[IPARM_AMALGAMATION] = 5;
iparm[IPARM_LEVEL_OF_FILL] = 0;
/* if (incomplete == 1)
{
dparm[DPARM_EPSILON_REFINEMENT] = 1e-7;
}
*/
/*
* Matrix needs :
* - to be in fortran numbering
* - to have only the lower triangular part in symmetric case
* - to have a graph with a symmetric structure in unsymmetric case
* If those criteria are not matched, the csc will be reallocated and changed.
*/
iparm[IPARM_MATRIX_VERIFICATION] = API_YES;
perm = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
invp = (pastix_int_t*)malloc(ncol*sizeof(pastix_int_t));
/*******************************************/
/* Step 1 - Ordering / Scotch */
/* Perform it only when the pattern of */
/* matrix change. */
/* eg: mesh refinement */
/* In many cases users can simply go from */
/* API_TASK_ORDERING to API_TASK_ANALYSE */
/* in one call. */
/*******************************************/
/*******************************************/
/* Step 2 - Symbolic factorization */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 3 - Mapping and Compute scheduling */
/* Perform it only when the pattern of */
/* matrix change. */
/*******************************************/
/*******************************************/
/* Step 4 - Numerical Factorisation */
/* Perform it each time the values of the */
/* matrix changed. */
/*******************************************/
iparm[IPARM_START_TASK] = API_TASK_ORDERING;
iparm[IPARM_END_TASK] = API_TASK_NUMFACT;
pastix(&pastix_data, MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, NULL, 1, iparm, dparm);
precond->int_array_1 = (magma_int_t*) perm;
precond->int_array_2 = (magma_int_t*) invp;
precond->M.val = (magmaFloatComplex*) values;
precond->M.col = (magma_int_t*) colptr;
precond->M.row = (magma_int_t*) rows;
precond->M.num_rows = A.num_rows;
precond->M.num_cols = A.num_cols;
precond->M.memory_location = Magma_CPU;
precond->pastix_data = pastix_data;
precond->iparm = iparm;
precond->dparm = dparm;
if( A.storage_type != Magma_CSR){
magma_c_mfree( &A_h1 );
}
magma_c_vfree( &b_h);
magma_c_mfree( &B );
#else
printf( "error: only real supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
magma_int_t
magma_capplypastix( magma_c_vector b, magma_c_vector *x,
magma_c_preconditioner *precond ){
#if defined(HAVE_PASTIX)
#if defined(PRECISION_d)
pastix_int_t ncol; /* Size of the matrix */
pastix_int_t *colptr = NULL; /* Indexes of first element of each column in row and values */
pastix_int_t *rows = NULL; /* Row of each element of the matrix */
pastix_float_t *values = NULL; /* Value of each element of the matrix */
pastix_float_t *rhs = NULL; /* right hand side */
pastix_int_t *iparm; /* integer parameters for pastix */
float *dparm; /* floating parameters for pastix */
pastix_int_t *perm = NULL; /* Permutation tabular */
pastix_int_t *invp = NULL; /* Reverse permutation tabular */
magma_c_vector b_h;
magma_c_vtransfer( b, &b_h, b.memory_location, Magma_CPU);
rhs = (pastix_float_t*) b_h.val;
ncol = precond->M.num_rows;
colptr = (pastix_int_t*) precond->M.col;
rows = (pastix_int_t*) precond->M.row;
values = (pastix_float_t*) precond->M.val;
iparm = precond->iparm;
dparm = precond->dparm;
perm = (pastix_int_t*)precond->int_array_1;
invp = (pastix_int_t*)precond->int_array_1;
/*******************************************/
/* Step 5 - Solve */
/* For each one of your Right-hand-side */
/* members. */
/* Also consider using multiple */
/* right-hand-side members. */
/*******************************************/
iparm[IPARM_START_TASK] = API_TASK_SOLVE;
iparm[IPARM_END_TASK] = API_TASK_REFINEMENT;
pastix(&(precond->pastix_data), MPI_COMM_WORLD,
ncol, colptr, rows, values,
perm, invp, b_h.val, 1, iparm, dparm);
// fix that x is not allocated every time
// in case of many iterations, it might be faster to use
// magma_csetvector( ncol,
// b_h.val, 1, x->val, 1 );
magma_c_vfree( x );
magma_c_vtransfer( b_h, x, Magma_CPU, b.memory_location);
magma_c_vfree( &b_h);
#else
printf( "error: only real supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_cgmres(
magma_c_sparse_matrix A,
magma_c_vector b,
magma_c_vector *x,
magma_c_solver_par *solver_par,
magma_queue_t queue )
{
magma_int_t stat = 0;
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
magma_int_t stat_cpu = 0, stat_dev = 0;
// prepare solver feedback
solver_par->solver = Magma_GMRES;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// local variables
magmaFloatComplex c_zero = MAGMA_C_ZERO, c_one = MAGMA_C_ONE,
c_mone = MAGMA_C_NEG_ONE;
magma_int_t dofs = A.num_rows;
magma_int_t i, j, k, m = 0;
magma_int_t restart = min( dofs-1, solver_par->restart );
magma_int_t ldh = restart+1;
float nom, rNorm, RNorm, nom0, betanom, r0 = 0.;
// CPU workspace
//magma_setdevice(0);
magmaFloatComplex *H, *HH, *y, *h1;
stat_cpu += magma_cmalloc_pinned( &H, (ldh+1)*ldh );
stat_cpu += magma_cmalloc_pinned( &y, ldh );
stat_cpu += magma_cmalloc_pinned( &HH, ldh*ldh );
stat_cpu += magma_cmalloc_pinned( &h1, ldh );
if( stat_cpu != 0){
magma_free_pinned( H );
magma_free_pinned( y );
magma_free_pinned( HH );
magma_free_pinned( h1 );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_HOST_ALLOC;
}
// GPU workspace
magma_c_vector r, q, q_t;
magma_c_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &q, Magma_DEV, dofs*(ldh+1), c_zero, queue );
q_t.memory_location = Magma_DEV;
q_t.dval = NULL;
q_t.num_rows = q_t.nnz = dofs; q_t.num_cols = 1;
magmaFloatComplex *dy = NULL, *dH = NULL;
stat_dev += magma_cmalloc( &dy, ldh );
stat_dev += magma_cmalloc( &dH, (ldh+1)*ldh );
if( stat_dev != 0){
magma_free_pinned( H );
magma_free_pinned( y );
magma_free_pinned( HH );
magma_free_pinned( h1 );
magma_free( dH );
magma_free( dy );
magma_free( dH );
magma_free( dy );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_DEVICE_ALLOC;
}
// GPU stream
magma_queue_t stream[2];
magma_event_t event[1];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
magma_event_create( &event[0] );
//magmablasSetKernelStream(stream[0]);
magma_cscal( dofs, c_zero, x->dval, 1 ); // x = 0
magma_ccopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
nom0 = betanom = magma_scnrm2( dofs, r.dval, 1 ); // nom0= || r||
nom = nom0 * nom0;
solver_par->init_res = nom0;
H(1,0) = MAGMA_C_MAKE( nom0, 0. );
magma_csetvector(1, &H(1,0), 1, &dH(1,0), 1);
if ( (r0 = nom0 * solver_par->epsilon ) < ATOLERANCE ){
r0 = solver_par->epsilon;
}
if ( nom < r0 ) {
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
for(k=1; k<=restart; k++) {
magma_ccopy(dofs, r.dval, 1, q(k-1), 1); // q[0] = 1.0/||r||
magma_cscal(dofs, 1./H(k,k-1), q(k-1), 1); // (to be fused)
q_t.dval = q(k-1);
//magmablasSetKernelStream(stream[0]);
magma_c_spmv( c_one, A, q_t, c_zero, r, queue ); // r = A q[k]
// if (solver_par->ortho == Magma_MGS ) {
// modified Gram-Schmidt
for (i=1; i<=k; i++) {
H(i,k) =magma_cdotc(dofs, q(i-1), 1, r.dval, 1);
// H(i,k) = q[i] . r
magma_caxpy(dofs,-H(i,k), q(i-1), 1, r.dval, 1);
// r = r - H(i,k) q[i]
}
H(k+1,k) = MAGMA_C_MAKE( magma_scnrm2(dofs, r.dval, 1), 0. ); // H(k+1,k) = ||r||
/*} else if (solver_par->ortho == Magma_FUSED_CGS ) {
// fusing cgemv with scnrm2 in classical Gram-Schmidt
magmablasSetKernelStream(stream[0]);
magma_ccopy(dofs, r.dval, 1, q(k), 1);
// dH(1:k+1,k) = q[0:k] . r
magmablas_cgemv(MagmaTrans, dofs, k+1, c_one, q(0),
dofs, r.dval, 1, c_zero, &dH(1,k), 1);
// r = r - q[0:k-1] dH(1:k,k)
magmablas_cgemv(MagmaNoTrans, dofs, k, c_mone, q(0),
dofs, &dH(1,k), 1, c_one, r.dval, 1);
// 1) dH(k+1,k) = sqrt( dH(k+1,k) - dH(1:k,k) )
magma_ccopyscale( dofs, k, r.dval, q(k), &dH(1,k) );
// 2) q[k] = q[k] / dH(k+1,k)
magma_event_record( event[0], stream[0] );
magma_queue_wait_event( stream[1], event[0] );
magma_cgetvector_async(k+1, &dH(1,k), 1, &H(1,k), 1, stream[1]);
// asynch copy dH(1:(k+1),k) to H(1:(k+1),k)
} else {
// classical Gram-Schmidt (default)
// > explicitly calling magmabls
magmablasSetKernelStream(stream[0]);
magmablas_cgemv(MagmaTrans, dofs, k, c_one, q(0),
dofs, r.dval, 1, c_zero, &dH(1,k), 1, queue );
// dH(1:k,k) = q[0:k-1] . r
#ifndef SCNRM2SCALE
// start copying dH(1:k,k) to H(1:k,k)
magma_event_record( event[0], stream[0] );
magma_queue_wait_event( stream[1], event[0] );
magma_cgetvector_async(k, &dH(1,k), 1, &H(1,k),
1, stream[1]);
#endif
// r = r - q[0:k-1] dH(1:k,k)
magmablas_cgemv(MagmaNoTrans, dofs, k, c_mone, q(0),
dofs, &dH(1,k), 1, c_one, r.dval, 1);
#ifdef SCNRM2SCALE
magma_ccopy(dofs, r.dval, 1, q(k), 1);
// q[k] = r / H(k,k-1)
magma_scnrm2scale(dofs, q(k), dofs, &dH(k+1,k) );
// dH(k+1,k) = sqrt(r . r) and r = r / dH(k+1,k)
magma_event_record( event[0], stream[0] );
// start sending dH(1:k,k) to H(1:k,k)
magma_queue_wait_event( stream[1], event[0] );
// can we keep H(k+1,k) on GPU and combine?
magma_cgetvector_async(k+1, &dH(1,k), 1, &H(1,k), 1, stream[1]);
#else
H(k+1,k) = MAGMA_C_MAKE( magma_scnrm2(dofs, r.dval, 1), 0. );
// H(k+1,k) = sqrt(r . r)
if ( k<solver_par->restart ) {
magmablasSetKernelStream(stream[0]);
magma_ccopy(dofs, r.dval, 1, q(k), 1);
// q[k] = 1.0/H[k][k-1] r
magma_cscal(dofs, 1./H(k+1,k), q(k), 1);
// (to be fused)
}
#endif
}*/
/* Minimization of || b-Ax || in H_k */
for (i=1; i<=k; i++) {
HH(k,i) = magma_cblas_cdotc( i+1, &H(1,k), 1, &H(1,i), 1 );
}
h1[k] = H(1,k)*H(1,0);
if (k != 1) {
for (i=1; i<k; i++) {
HH(k,i) = HH(k,i)/HH(i,i);//
for (m=i+1; m<=k; m++) {
HH(k,m) -= HH(k,i) * HH(m,i) * HH(i,i);
}
h1[k] -= h1[i] * HH(k,i);
}
}
y[k] = h1[k]/HH(k,k);
if (k != 1)
for (i=k-1; i>=1; i--) {
y[i] = h1[i]/HH(i,i);
for (j=i+1; j<=k; j++)
y[i] -= y[j] * HH(j,i);
}
m = k;
rNorm = fabs(MAGMA_C_REAL(H(k+1,k)));
}/* Minimization done */
// compute solution approximation
magma_csetmatrix(m, 1, y+1, m, dy, m );
magma_cgemv(MagmaNoTrans, dofs, m, c_one, q(0), dofs, dy, 1,
c_one, x->dval, 1);
// compute residual
magma_c_spmv( c_mone, A, *x, c_zero, r, queue ); // r = - A * x
magma_caxpy(dofs, c_one, b.dval, 1, r.dval, 1); // r = r + b
H(1,0) = MAGMA_C_MAKE( magma_scnrm2(dofs, r.dval, 1), 0. );
// RNorm = H[1][0] = || r ||
RNorm = MAGMA_C_REAL( H(1,0) );
betanom = fabs(RNorm);
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < r0 ) {
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_cresidual( A, b, *x, &residual, queue );
solver_par->iter_res = betanom;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter) {
solver_par->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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_SLOW_CONVERGENCE;
}
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
// free pinned memory
magma_free_pinned( H );
magma_free_pinned( y );
magma_free_pinned( HH );
magma_free_pinned( h1 );
// free GPU memory
magma_free(dy);
if (dH != NULL ) magma_free(dH);
magma_c_vfree(&r, queue );
magma_c_vfree(&q, queue );
// free GPU streams and events
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_event_destroy( event[0] );
//magmablasSetKernelStream(NULL);
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_cgmres */
extern "C" magma_int_t
magma_cbicgstab_merge(
magma_c_sparse_matrix A, magma_c_vector b,
magma_c_vector *x, magma_c_solver_par *solver_par,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
// prepare solver feedback
solver_par->solver = Magma_BICGSTABMERGE;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// some useful variables
magmaFloatComplex c_zero = MAGMA_C_ZERO, c_one = MAGMA_C_ONE;
magma_int_t dofs = A.num_rows;
// GPU stream
magma_queue_t stream[2];
magma_event_t event[1];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
magma_event_create( &event[0] );
// workspace
magma_c_vector q, r,rr,p,v,s,t;
magmaFloatComplex *d1, *d2, *skp;
d1 = NULL;
d2 = NULL;
skp = NULL;
magma_int_t stat_dev = 0, stat_cpu = 0;
stat_dev += magma_cmalloc( &d1, dofs*(2) );
stat_dev += magma_cmalloc( &d2, dofs*(2) );
// array for the parameters
stat_dev += magma_cmalloc( &skp, 8 );
if( stat_dev != 0 ){
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
printf("error: memory allocation.\n");
return MAGMA_ERR_DEVICE_ALLOC;
}
// skp = [alpha|beta|omega|rho_old|rho|nom|tmp1|tmp2]
magma_c_vinit( &q, Magma_DEV, dofs*6, c_zero, queue );
// q = rr|r|p|v|s|t
rr.memory_location = Magma_DEV; rr.dval = NULL; rr.num_rows = rr.nnz = dofs; rr.num_cols = 1;
r.memory_location = Magma_DEV; r.dval = NULL; r.num_rows = r.nnz = dofs; r.num_cols = 1;
p.memory_location = Magma_DEV; p.dval = NULL; p.num_rows = p.nnz = dofs; p.num_cols = 1;
v.memory_location = Magma_DEV; v.dval = NULL; v.num_rows = v.nnz = dofs; v.num_cols = 1;
s.memory_location = Magma_DEV; s.dval = NULL; s.num_rows = s.nnz = dofs; s.num_cols = 1;
t.memory_location = Magma_DEV; t.dval = NULL; t.num_rows = t.nnz = dofs; t.num_cols = 1;
rr.dval = q(0);
r.dval = q(1);
p.dval = q(2);
v.dval = q(3);
s.dval = q(4);
t.dval = q(5);
// solver variables
magmaFloatComplex alpha, beta, omega, rho_old, rho_new, *skp_h;
float nom, nom0, betanom, r0, den;
// solver setup
magma_cscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_ccopy( dofs, b.dval, 1, q(0), 1 ); // rr = b
magma_ccopy( dofs, b.dval, 1, q(1), 1 ); // r = b
rho_new = magma_cdotc( dofs, r.dval, 1, r.dval, 1 ); // rho=<rr,r>
nom = MAGMA_C_REAL(magma_cdotc( dofs, r.dval, 1, r.dval, 1 ));
nom0 = betanom = sqrt(nom); // nom = || r ||
rho_old = omega = alpha = MAGMA_C_MAKE( 1.0, 0. );
beta = rho_new;
solver_par->init_res = nom0;
// array on host for the parameters
stat_cpu = magma_cmalloc_cpu( &skp_h, 8 );
if( stat_cpu != 0 ){
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
magma_free_cpu( skp_h );
printf("error: memory allocation.\n");
return MAGMA_ERR_HOST_ALLOC;
}
skp_h[0]=alpha;
skp_h[1]=beta;
skp_h[2]=omega;
skp_h[3]=rho_old;
skp_h[4]=rho_new;
skp_h[5]=MAGMA_C_MAKE(nom, 0.0);
magma_csetvector( 8, skp_h, 1, skp, 1 );
magma_c_spmv( c_one, A, r, c_zero, v, queue ); // z = A r
den = MAGMA_C_REAL( magma_cdotc(dofs, v.dval, 1, r.dval, 1) );// den = z dot r
if ( (r0 = nom * solver_par->epsilon) < ATOLERANCE )
r0 = ATOLERANCE;
if ( nom < r0 ) {
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
// check positive definite
if (den <= 0.0) {
printf("Operator A is not postive definite. (Ar,r) = %f\n", den);
magmablasSetKernelStream( orig_queue );
return MAGMA_NONSPD;
solver_par->info = MAGMA_NONSPD;;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
magmablasSetKernelStream(stream[0]);
// computes p=r+beta*(p-omega*v)
magma_cbicgmerge1( dofs, skp, v.dval, r.dval, p.dval, queue );
magma_c_spmv( c_one, A, p, c_zero, v, queue ); // v = Ap
magma_cmdotc( dofs, 1, q.dval, v.dval, d1, d2, skp, queue );
magma_cbicgmerge4( 1, skp, queue );
magma_cbicgmerge2( dofs, skp, r.dval, v.dval, s.dval, queue ); // s=r-alpha*v
magma_c_spmv( c_one, A, s, c_zero, t, queue ); // t=As
magma_cmdotc( dofs, 2, q.dval+4*dofs, t.dval, d1, d2, skp+6, queue );
magma_cbicgmerge4( 2, skp, queue );
magma_cbicgmerge_xrbeta( dofs, d1, d2, q.dval, r.dval, p.dval,
s.dval, t.dval, x->dval, skp, queue );
// check stopping criterion (asynchronous copy)
magma_cgetvector_async( 1 , skp+5, 1,
skp_h+5, 1, stream[1] );
betanom = sqrt(MAGMA_C_REAL(skp_h[5]));
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < r0 ) {
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_cresidual( A, b, *x, &residual, queue );
solver_par->iter_res = betanom;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter) {
solver_par->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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_SLOW_CONVERGENCE;
}
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
magma_c_vfree(&q, queue ); // frees all vectors
magma_free(d1);
magma_free(d2);
magma_free( skp );
magma_free_cpu( skp_h );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* cbicgstab_merge */
/* ////////////////////////////////////////////////////////////////////////////
-- testing sparse matrix vector product
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magma_queue_t queue;
magma_queue_create( /*devices[ opts->device ],*/ &queue );
magma_c_sparse_matrix hA, hA_SELLP, hA_ELL, dA, dA_SELLP, dA_ELL;
hA_SELLP.blocksize = 8;
hA_SELLP.alignment = 8;
real_Double_t start, end, res;
magma_int_t *pntre;
magmaFloatComplex c_one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex c_zero = MAGMA_C_MAKE(0.0, 0.0);
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_cspmv"
" [ --blocksize %d --alignment %d (for SELLP) ]"
" matrices \n\n", (int) hA_SELLP.blocksize, (int) 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] );
magma_cm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_c_csr_mtx( &hA, argv[i], queue );
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
(int) hA.num_rows,(int) hA.num_cols,(int) hA.nnz );
real_Double_t FLOPS = 2.0*hA.nnz/1e9;
magma_c_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_c_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_c_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_c_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
#ifdef MAGMA_WITH_MKL
// calling MKL with CSR
pntre = (magma_int_t*)malloc( (hA.num_rows+1)*sizeof(magma_int_t) );
pntre[0] = 0;
for (j=0; j<hA.num_rows; 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 *col;
TESTING_MALLOC_CPU( col, MKL_INT, nnz );
for( magma_int_t t=0; t < hA.nnz; ++t ) {
col[ t ] = hA.col[ t ];
}
MKL_INT *row;
TESTING_MALLOC_CPU( row, MKL_INT, num_rows );
for( magma_int_t t=0; t < hA.num_rows; ++t ) {
row[ t ] = hA.col[ t ];
}
start = magma_wtime();
for (j=0; j<10; j++ ) {
mkl_ccsrmv( "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 : %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
TESTING_FREE_CPU( row );
TESTING_FREE_CPU( col );
free(pntre);
#endif // MAGMA_WITH_MKL
// copy matrix to GPU
magma_c_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
// SpMV on GPU (CSR) -- this is the reference!
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_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/(end-start) );
magma_c_mfree(&dA, queue );
magma_c_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_c_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_c_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_c_mfree(&hA_ELL, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (ELL)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_spmv( c_one, dA_ELL, dx, c_zero, dy, queue );
end = magma_sync_wtime( queue );
printf( " > MAGMA: %.2e seconds %.2e GFLOP/s (standard ELL).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_mfree(&dA_ELL, queue );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_c_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_c_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_c_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_c_mfree(&hA_SELLP, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// SpMV on GPU (SELLP)
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
magma_c_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/(end-start) );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv SELL-P: ok\n");
else
printf("# tester spmv SELL-P: failed\n");
magma_c_vfree( &hcheck, queue );
magma_c_mfree(&dA_SELLP, queue );
// SpMV on GPU (CUSPARSE - CSR)
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
magmaFloatComplex alpha = c_one;
magmaFloatComplex beta = c_zero;
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_c_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseCcsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
hA.num_rows, hA.num_cols, hA.nnz, &alpha, descr,
dA.dval, dA.drow, dA.dcol, dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE CSR\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (CSR).\n",
(end-start)/10, FLOPS*10/(end-start) );
cusparseMatDescr_t descrA;
cusparseStatus = cusparseCreateMatDescr(&descrA);
if (cusparseStatus != 0) printf("error\n");
cusparseHybMat_t hybA;
cusparseStatus = cusparseCreateHybMat( &hybA );
if (cusparseStatus != 0) printf("error\n");
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE CSR: ok\n");
else
printf("# tester spmv cuSPARSE CSR: failed\n");
magma_c_vfree( &hcheck, queue );
magma_c_vfree( &dy, queue );
magma_c_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseCcsr2hyb(cusparseHandle, hA.num_rows, hA.num_cols,
descrA, dA.dval, dA.drow, dA.dcol,
hybA, 0, CUSPARSE_HYB_PARTITION_AUTO);
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseChybmv( cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
&alpha, descrA, hybA,
dx.dval, &beta, dy.dval);
end = magma_sync_wtime( queue );
if (cusparseStatus != 0) printf("error in cuSPARSE HYB\n");
printf( " > CUSPARSE: %.2e seconds %.2e GFLOP/s (HYB).\n",
(end-start)/10, FLOPS*10/(end-start) );
magma_c_vtransfer( dy, &hcheck , Magma_DEV, Magma_CPU, queue );
res = 0.0;
for(magma_int_t k=0; k<hA.num_rows; k++ )
res=res + MAGMA_C_REAL(hcheck.val[k]) - MAGMA_C_REAL(hrefvec.val[k]);
printf("# |x-y|_F = %8.2e\n", res);
if ( res < .000001 )
printf("# tester spmv cuSPARSE HYB: ok\n");
else
printf("# tester spmv cuSPARSE HYB: failed\n");
magma_c_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_c_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_c_mfree(&hA, queue );
magma_c_vfree(&hx, queue );
magma_c_vfree(&hy, queue );
magma_c_vfree(&hrefvec, queue );
// free GPU memory
magma_c_vfree(&dx, queue );
magma_c_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_cbicgstab(
magma_c_sparse_matrix A, magma_c_vector b, magma_c_vector *x,
magma_c_solver_par *solver_par,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
// prepare solver feedback
solver_par->solver = Magma_BICGSTAB;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// some useful variables
magmaFloatComplex c_zero = MAGMA_C_ZERO, c_one = MAGMA_C_ONE,
c_mone = MAGMA_C_NEG_ONE;
magma_int_t dofs = A.num_rows;
// workspace
magma_c_vector r,rr,p,v,s,t;
magma_c_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &rr, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &p, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &v, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &s, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &t, Magma_DEV, dofs, c_zero, queue );
// solver variables
magmaFloatComplex alpha, beta, omega, rho_old, rho_new;
float nom, betanom, nom0, r0, den, res;
// solver setup
magma_cscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_ccopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
magma_ccopy( dofs, b.dval, 1, rr.dval, 1 ); // rr = b
nom0 = betanom = magma_scnrm2( dofs, r.dval, 1 ); // nom = || r ||
nom = nom0*nom0;
rho_old = omega = alpha = MAGMA_C_MAKE( 1.0, 0. );
solver_par->init_res = nom0;
magma_c_spmv( c_one, A, r, c_zero, v, queue ); // z = A r
den = MAGMA_C_REAL( magma_cdotc(dofs, v.dval, 1, r.dval, 1) ); // den = z' * r
if ( (r0 = nom * solver_par->epsilon) < ATOLERANCE )
r0 = ATOLERANCE;
if ( nom < r0 ) {
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
// check positive definite
if (den <= 0.0) {
printf("Operator A is not postive definite. (Ar,r) = %f\n", den);
magmablasSetKernelStream( orig_queue );
return MAGMA_NONSPD;
solver_par->info = MAGMA_NONSPD;;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
rho_new = magma_cdotc( dofs, rr.dval, 1, r.dval, 1 ); // rho=<rr,r>
beta = rho_new/rho_old * alpha/omega; // beta=rho/rho_old *alpha/omega
magma_cscal( dofs, beta, p.dval, 1 ); // p = beta*p
magma_caxpy( dofs, c_mone * omega * beta, v.dval, 1 , p.dval, 1 );
// p = p-omega*beta*v
magma_caxpy( dofs, c_one, r.dval, 1, p.dval, 1 ); // p = p+r
magma_c_spmv( c_one, A, p, c_zero, v, queue ); // v = Ap
alpha = rho_new / magma_cdotc( dofs, rr.dval, 1, v.dval, 1 );
magma_ccopy( dofs, r.dval, 1 , s.dval, 1 ); // s=r
magma_caxpy( dofs, c_mone * alpha, v.dval, 1 , s.dval, 1 ); // s=s-alpha*v
magma_c_spmv( c_one, A, s, c_zero, t, queue ); // t=As
omega = magma_cdotc( dofs, t.dval, 1, s.dval, 1 ) // omega = <s,t>/<t,t>
/ magma_cdotc( dofs, t.dval, 1, t.dval, 1 );
magma_caxpy( dofs, alpha, p.dval, 1 , x->dval, 1 ); // x=x+alpha*p
magma_caxpy( dofs, omega, s.dval, 1 , x->dval, 1 ); // x=x+omega*s
magma_ccopy( dofs, s.dval, 1 , r.dval, 1 ); // r=s
magma_caxpy( dofs, c_mone * omega, t.dval, 1 , r.dval, 1 ); // r=r-omega*t
res = betanom = magma_scnrm2( dofs, r.dval, 1 );
nom = betanom*betanom;
rho_old = rho_new; // rho_old=rho
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/nom0 < solver_par->epsilon ) {
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_cresidual( A, b, *x, &residual, queue );
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter) {
solver_par->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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_SLOW_CONVERGENCE;
}
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
magma_c_vfree(&r, queue );
magma_c_vfree(&rr, queue );
magma_c_vfree(&p, queue );
magma_c_vfree(&v, queue );
magma_c_vfree(&s, queue );
magma_c_vfree(&t, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_cbicgstab */
extern "C" magma_int_t
magma_cbpcg(
magma_c_sparse_matrix A, magma_c_vector b, magma_c_vector *x,
magma_c_solver_par *solver_par,
magma_c_preconditioner *precond_par,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
magma_int_t stat_dev = 0, stat_cpu = 0;
magma_int_t i, num_vecs = b.num_rows/A.num_rows;
// prepare solver feedback
solver_par->solver = Magma_PCG;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// local variables
magmaFloatComplex c_zero = MAGMA_C_ZERO, c_one = MAGMA_C_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_c_vector r, rt, p, q, h;
magma_c_vinit( &r, Magma_DEV, dofs*num_vecs, c_zero, queue );
magma_c_vinit( &rt, Magma_DEV, dofs*num_vecs, c_zero, queue );
magma_c_vinit( &p, Magma_DEV, dofs*num_vecs, c_zero, queue );
magma_c_vinit( &q, Magma_DEV, dofs*num_vecs, c_zero, queue );
magma_c_vinit( &h, Magma_DEV, dofs*num_vecs, c_zero, queue );
// solver variables
magmaFloatComplex *alpha, *beta;
alpha = NULL;
beta = NULL;
stat_cpu += magma_cmalloc_cpu(&alpha, num_vecs);
stat_cpu += magma_cmalloc_cpu(&beta, num_vecs);
float *nom, *nom0, *r0, *gammaold, *gammanew, *den, *res, *residual;
nom = NULL;
nom0 = NULL;
r0 = NULL;
gammaold = NULL;
gammanew = NULL;
den = NULL;
res = NULL;
residual = NULL;
stat_cpu += magma_smalloc_cpu(&residual, num_vecs);
stat_cpu += magma_smalloc_cpu(&nom, num_vecs);
stat_cpu += magma_smalloc_cpu(&nom0, num_vecs);
stat_cpu += magma_smalloc_cpu(&r0, num_vecs);
stat_cpu += magma_smalloc_cpu(&gammaold, num_vecs);
stat_cpu += magma_smalloc_cpu(&gammanew, num_vecs);
stat_cpu += magma_smalloc_cpu(&den, num_vecs);
stat_cpu += magma_smalloc_cpu(&res, num_vecs);
stat_cpu += magma_smalloc_cpu(&residual, num_vecs);
if( stat_cpu != 0 ){
magma_free_cpu( nom );
magma_free_cpu( nom0 );
magma_free_cpu( r0 );
magma_free_cpu( gammaold );
magma_free_cpu( gammanew );
magma_free_cpu( den );
magma_free_cpu( res );
magma_free_cpu( alpha );
magma_free_cpu( beta );
magma_free_cpu( residual );
magmablasSetKernelStream( orig_queue );
printf("error: memory allocation.\n");
return MAGMA_ERR_HOST_ALLOC;
}
// solver setup
magma_cscal( dofs*num_vecs, c_zero, x->dval, 1) ; // x = 0
magma_ccopy( dofs*num_vecs, b.dval, 1, r.dval, 1 ); // r = b
// preconditioner
magma_c_applyprecond_left( A, r, &rt, precond_par, queue );
magma_c_applyprecond_right( A, rt, &h, precond_par, queue );
magma_ccopy( dofs*num_vecs, h.dval, 1, p.dval, 1 ); // p = h
for( i=0; i<num_vecs; i++) {
nom[i] = MAGMA_C_REAL( magma_cdotc(dofs, r(i), 1, h(i), 1) );
nom0[i] = magma_scnrm2( dofs, r(i), 1 );
}
magma_c_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
for( i=0; i<num_vecs; i++)
den[i] = MAGMA_C_REAL( magma_cdotc(dofs, p(i), 1, q(i), 1) ); // den = p dot q
solver_par->init_res = nom0[0];
if ( (r0[0] = nom[0] * solver_par->epsilon) < ATOLERANCE )
r0[0] = ATOLERANCE;
// check positive definite
if (den[0] <= 0.0) {
printf("Operator A is not postive definite. (Ar,r) = %f\n", den[0]);
magmablasSetKernelStream( orig_queue );
return MAGMA_NONSPD;
solver_par->info = MAGMA_NONSPD;;
}
if ( nom[0] < r0[0] ) {
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nom0[0];
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
// preconditioner
magma_c_applyprecond_left( A, r, &rt, precond_par, queue );
magma_c_applyprecond_right( A, rt, &h, precond_par, queue );
for( i=0; i<num_vecs; i++)
gammanew[i] = MAGMA_C_REAL( magma_cdotc(dofs, r(i), 1, h(i), 1) ); // gn = < r,h>
if ( solver_par->numiter==1 ) {
magma_ccopy( dofs*num_vecs, h.dval, 1, p.dval, 1 ); // p = h
} else {
for( i=0; i<num_vecs; i++) {
beta[i] = MAGMA_C_MAKE(gammanew[i]/gammaold[i], 0.); // beta = gn/go
magma_cscal(dofs, beta[i], p(i), 1); // p = beta*p
magma_caxpy(dofs, c_one, h(i), 1, p(i), 1); // p = p + h
}
}
magma_c_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
// magma_c_bspmv_tuned( dofs, num_vecs, c_one, A, p.dval, c_zero, q.dval, queue );
for( i=0; i<num_vecs; i++) {
den[i] = MAGMA_C_REAL(magma_cdotc(dofs, p(i), 1, q(i), 1));
// den = p dot q
alpha[i] = MAGMA_C_MAKE(gammanew[i]/den[i], 0.);
magma_caxpy(dofs, alpha[i], p(i), 1, x->dval+dofs*i, 1); // x = x + alpha p
magma_caxpy(dofs, -alpha[i], q(i), 1, r(i), 1); // r = r - alpha q
gammaold[i] = gammanew[i];
res[i] = magma_scnrm2( dofs, r(i), 1 );
}
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[0];
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res[0]/nom0[0] < solver_par->epsilon ) {
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
magma_cresidual( A, b, *x, residual, queue );
solver_par->iter_res = res[0];
solver_par->final_res = residual[0];
if ( solver_par->numiter < solver_par->maxiter) {
solver_par->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[0];
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_SLOW_CONVERGENCE;
}
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[0];
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
for( i=0; i<num_vecs; i++) {
printf("%.4e ",res[i]);
}
printf("\n");
for( i=0; i<num_vecs; i++) {
printf("%.4e ",residual[i]);
}
printf("\n");
magma_c_vfree(&r, queue );
magma_c_vfree(&rt, queue );
magma_c_vfree(&p, queue );
magma_c_vfree(&q, queue );
magma_c_vfree(&h, queue );
magma_free_cpu(alpha);
magma_free_cpu(beta);
magma_free_cpu(nom);
magma_free_cpu(nom0);
magma_free_cpu(r0);
magma_free_cpu(gammaold);
magma_free_cpu(gammanew);
magma_free_cpu(den);
magma_free_cpu(res);
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_cbpcg */
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_copts zopts;
int i=1;
magma_cparse_opts( argc, argv, &zopts, &i);
magmaFloatComplex one = MAGMA_C_MAKE(1.0, 0.0);
magmaFloatComplex zero = MAGMA_C_MAKE(0.0, 0.0);
magma_c_sparse_matrix A, B, B_d;
magma_c_vector x, b;
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.precond_par.solver = Magma_NONE;
magma_csolverinfo_init( &zopts.solver_par, &zopts.precond_par );
while( i < argc ){
magma_c_csr_mtx( &A, argv[i] );
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
magma_cmscale( &A, zopts.scaling );
magma_c_mconvert( A, &B, Magma_CSR, zopts.output_format );
magma_c_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_c_vinit( &b, Magma_DEV, A.num_cols, one );
magma_c_vinit( &x, Magma_DEV, A.num_cols, one );
magma_c_spmv( one, B_d, x, zero, b ); // b = A x
magma_c_vfree(&x);
magma_c_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_c_solver( B_d, b, &x, &zopts );
magma_csolverinfo( &zopts.solver_par, &zopts.precond_par );
magma_c_mfree(&B_d);
magma_c_mfree(&B);
magma_c_mfree(&A);
magma_c_vfree(&x);
magma_c_vfree(&b);
i++;
}
magma_csolverinfo_free( &zopts.solver_par, &zopts.precond_par );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_ccg_res(
magma_c_sparse_matrix A, magma_c_vector b, magma_c_vector *x,
magma_c_solver_par *solver_par,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
// prepare solver feedback
solver_par->solver = Magma_CG;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// local variables
magmaFloatComplex c_zero = MAGMA_C_ZERO, c_one = MAGMA_C_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_c_vector r, p, q;
magma_c_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &p, Magma_DEV, dofs, c_zero, queue );
magma_c_vinit( &q, Magma_DEV, dofs, c_zero, queue );
// solver variables
magmaFloatComplex alpha, beta;
float nom, nom0, r0, betanom, betanomsq, den, res;
// solver setup
magma_cscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_ccopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
magma_ccopy( dofs, b.dval, 1, p.dval, 1 ); // p = b
nom0 = betanom = magma_scnrm2( dofs, r.dval, 1 );
nom = nom0 * nom0; // nom = r' * r
magma_c_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_C_REAL( magma_cdotc(dofs, p.dval, 1, q.dval, 1) );// den = p dot q
solver_par->init_res = nom0;
if ( (r0 = nom * solver_par->epsilon) < ATOLERANCE )
r0 = ATOLERANCE;
if ( nom < r0 ) {
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
// check positive definite
if (den <= 0.0) {
printf("Operator A is not postive definite. (Ar,r) = %f\n", den);
magmablasSetKernelStream( orig_queue );
return MAGMA_NONSPD;
solver_par->info = MAGMA_NONSPD;;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nom0;
solver_par->timing[0] = 0.0;
}
// start iteration
for( solver_par->numiter= 1; solver_par->numiter<solver_par->maxiter;
solver_par->numiter++ ) {
alpha = MAGMA_C_MAKE(nom/den, 0.);
magma_caxpy(dofs, alpha, p.dval, 1, x->dval, 1); // x = x + alpha p
magma_caxpy(dofs, -alpha, q.dval, 1, r.dval, 1); // r = r - alpha q
res = betanom = magma_scnrm2(dofs, r.dval, 1); // betanom = || r ||
betanomsq = betanom * betanom; // betanoms = r' * r
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/nom0 < solver_par->epsilon ) {
break;
}
beta = MAGMA_C_MAKE(betanomsq/nom, 0.); // beta = betanoms/nom
magma_cscal(dofs, beta, p.dval, 1); // p = beta*p
magma_caxpy(dofs, c_one, r.dval, 1, p.dval, 1); // p = p + r
magma_c_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_C_REAL(magma_cdotc(dofs, p.dval, 1, q.dval, 1));
// den = p dot q
nom = betanomsq;
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_cresidual( A, b, *x, &residual, queue );
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter) {
solver_par->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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_SLOW_CONVERGENCE;
}
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
magma_c_vfree(&r, queue );
magma_c_vfree(&p, queue );
magma_c_vfree(&q, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_ccg */