extern "C" magma_int_t
magma_zjacobiiter(
magma_z_sparse_matrix M, magma_z_vector c, magma_z_vector *x,
magma_z_solver_par *solver_par,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
// local variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE,
c_mone = MAGMA_Z_NEG_ONE;
magma_int_t dofs = M.num_rows;
magma_z_vector t, swap;
magma_z_vinit( &t, Magma_DEV, dofs, c_zero, queue );
for( magma_int_t i=0; i<solver_par->maxiter; i++ ) {
magma_z_spmv( c_mone, M, *x, c_zero, t, queue ); // t = - M * x
magma_zaxpy( dofs, c_one , c.dval, 1 , t.dval, 1 ); // t = t + c
// swap so that x again contains solution, and y is ready to be used
swap = *x;
*x = t;
t = swap;
//magma_zcopy( dofs, t.dval, 1 , x->dval, 1 ); // x = t
}
magma_z_vfree( &t, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_zjacobiiter */
extern "C" magma_int_t
magma_z_applyprecond(
magma_z_sparse_matrix A,
magma_z_vector b,
magma_z_vector *x,
magma_z_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_zjacobi_diagscal( A.num_rows, precond->d, b, x, queue );
}
else if ( precond->solver == Magma_PASTIX ) {
magma_zapplypastix( b, x, precond, queue );
}
else if ( precond->solver == Magma_ILU ) {
magma_z_vector tmp;
magma_z_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vfree( &tmp, queue );
}
else if ( precond->solver == Magma_ICC ) {
magma_z_vector tmp;
magma_z_vinit( &tmp, Magma_DEV, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vfree( &tmp, queue );
}
else if ( precond->solver == Magma_NONE ) {
magma_zcopy( 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;
}
extern "C" magma_int_t
magma_zjacobisetup_vector(
magma_z_vector b, magma_z_vector d,
magma_z_vector *c,
magma_queue_t queue )
{
if ( b.memory_location == Magma_CPU ) {
magma_z_vector diag, c_t, b_h;
magma_z_vinit( &c_t, Magma_CPU, b.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vtransfer( b, &b_h, b.memory_location, Magma_CPU, queue );
magma_z_vtransfer( 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];
}
magma_z_vtransfer( c_t, c, Magma_CPU, b.memory_location, queue );
magma_z_vfree( &diag, queue );
magma_z_vfree( &c_t, queue );
magma_z_vfree( &b_h, queue );
return MAGMA_SUCCESS;
}
else if ( b.memory_location == Magma_DEV ) {
// fill vector
magma_z_vector tmp;
magma_z_vinit( &tmp, Magma_DEV, b.num_rows, MAGMA_Z_ZERO, queue );
magma_zjacobisetup_vector_gpu(
b.num_rows, b, d, *c, &tmp, queue );
magma_z_vfree( &tmp, queue );
return MAGMA_SUCCESS;
}
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_zjacobisetup_diagscal(
magma_z_sparse_matrix A, magma_z_vector *d,
magma_queue_t queue )
{
magma_int_t i;
magma_z_sparse_matrix A_h1, B;
magma_z_vector diag;
magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue );
if ( A.storage_type != Magma_CSR) {
magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue );
magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue );
}
else {
magma_z_mtransfer( 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] = 1.0/B.val[i];
if ( MAGMA_Z_REAL( diag.val[rowindex]) == 0 )
printf(" error: zero diagonal element in row %d!\n",
(int) rowindex);
}
}
}
magma_z_vtransfer( diag, d, Magma_CPU, A.memory_location, queue );
if ( A.storage_type != Magma_CSR) {
magma_z_mfree( &A_h1, queue );
}
magma_z_mfree( &B, queue );
magma_z_vfree( &diag, queue );
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_ziterref(
magma_z_sparse_matrix A, magma_z_vector b, magma_z_vector *x,
magma_z_solver_par *solver_par, magma_z_preconditioner *precond_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_ITERREF;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
double residual;
magma_zresidual( A, b, *x, &residual, queue );
solver_par->init_res = residual;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE,
c_mone = MAGMA_Z_NEG_ONE;
magma_int_t dofs = A.num_rows;
// workspace
magma_z_vector r,z;
magma_z_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_z_vinit( &z, Magma_DEV, dofs, c_zero, queue );
// solver variables
double nom, nom0, r0;
// solver setup
magma_zscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_zcopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
nom0 = magma_dznrm2(dofs, r.dval, 1); // nom0 = || r ||
nom = nom0 * nom0;
solver_par->init_res = nom0;
if ( (r0 = nom * solver_par->epsilon) < ATOLERANCE )
r0 = ATOLERANCE;
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++ ) {
magma_zscal( dofs, MAGMA_Z_MAKE(1./nom, 0.), r.dval, 1) ; // scale it
magma_z_precond( A, r, &z, precond_par, queue ); // inner solver: A * z = r
magma_zscal( dofs, MAGMA_Z_MAKE(nom, 0.), z.dval, 1) ; // scale it
magma_zaxpy(dofs, c_one, z.dval, 1, x->dval, 1); // x = x + z
magma_z_spmv( c_mone, A, *x, c_zero, r, queue ); // r = - A x
magma_zaxpy(dofs, c_one, b.dval, 1, r.dval, 1); // r = r + b
nom = magma_dznrm2(dofs, r.dval, 1); // nom = || 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) nom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( nom < r0 ) {
break;
}
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
magma_zresidual( A, b, *x, &residual, queue );
solver_par->final_res = residual;
solver_par->iter_res = nom;
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) nom;
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) nom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
magma_z_vfree(&r, queue );
magma_z_vfree(&z, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_ziterref */
extern "C" magma_int_t
magma_zpastixsetup(
magma_z_sparse_matrix A, magma_z_vector b,
magma_z_preconditioner *precond,
magma_queue_t queue )
{
#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 */
double *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_z_sparse_matrix A_h1, B;
magma_z_vector diag, c_t, b_h;
magma_z_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vtransfer( b, &b_h, A.memory_location, Magma_CPU, queue );
if ( A.storage_type != Magma_CSR ) {
magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue );
magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue );
}
else {
magma_z_mtransfer( A, &B, A.memory_location, Magma_CPU, queue );
}
rhs = (pastix_float_t*) b_h.dval;
ncol = B.num_rows;
colptr = B.drow;
rows = B.dcol;
values = (pastix_float_t*) B.dval;
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.dval = (magmaDoubleComplex*) values;
precond->M.dcol = (magma_int_t*) colptr;
precond->M.drow = (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_z_mfree( &A_h1, queue );
}
magma_z_vfree( &b_h, queue );
magma_z_mfree( &B, queue );
#else
printf( "error: only double precision supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_zjacobi(
magma_z_sparse_matrix A,
magma_z_vector b,
magma_z_vector *x,
magma_z_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_JACOBI;
solver_par->info = MAGMA_SUCCESS;
real_Double_t tempo1, tempo2;
double residual;
magma_zresidual( A, b, *x, &residual, queue );
solver_par->init_res = residual;
solver_par->res_vec = NULL;
solver_par->timing = NULL;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE,
c_mone = MAGMA_Z_NEG_ONE;
magma_int_t dofs = A.num_rows;
double nom0;
magma_z_sparse_matrix M;
magma_z_vector c, r;
magma_z_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_z_spmv( c_one, A, *x, c_zero, r, queue ); // r = A x
magma_zaxpy(dofs, c_mone, b.dval, 1, r.dval, 1); // r = r - b
nom0 = magma_dznrm2(dofs, r.dval, 1); // den = || r ||
// Jacobi setup
magma_zjacobisetup( A, b, &M, &c, queue );
magma_z_solver_par jacobiiter_par;
jacobiiter_par.maxiter = solver_par->maxiter;
tempo1 = magma_sync_wtime( queue );
// Jacobi iterator
magma_zjacobiiter( M, c, x, &jacobiiter_par, queue );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
magma_zresidual( A, b, *x, &residual, queue );
solver_par->final_res = residual;
solver_par->numiter = solver_par->maxiter;
if ( solver_par->init_res > solver_par->final_res )
solver_par->info = MAGMA_SUCCESS;
else
solver_par->info = MAGMA_DIVERGENCE;
magma_z_mfree( &M, queue );
magma_z_vfree( &c, queue );
magma_z_vfree( &r, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_zjacobi */
extern "C" magma_int_t
magma_zjacobisetup(
magma_z_sparse_matrix A, magma_z_vector b,
magma_z_sparse_matrix *M, magma_z_vector *c,
magma_queue_t queue )
{
magma_int_t i;
magma_z_sparse_matrix A_h1, A_h2, B, C;
magma_z_vector diag, c_t, b_h;
magma_z_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_Z_ZERO, queue );
magma_z_vtransfer( b, &b_h, A.memory_location, Magma_CPU, queue );
if ( A.storage_type != Magma_CSR ) {
magma_z_mtransfer( A, &A_h1, A.memory_location, Magma_CPU, queue );
magma_z_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR, queue );
}
else {
magma_z_mtransfer( 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_Z_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_Z_MAKE( 0., 0. );
}
}
c_t.val[rowindex] = b_h.val[rowindex] / diag.val[rowindex];
}
magma_z_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;
magma_z_mconvert( C, &A_h2, Magma_CSR, A_h1.storage_type, queue );
magma_z_mtransfer( A_h2, M, Magma_CPU, A.memory_location, queue );
}
else {
magma_z_mtransfer( C, M, Magma_CPU, A.memory_location, queue );
}
magma_z_vtransfer( c_t, c, Magma_CPU, A.memory_location, queue );
if ( A.storage_type != Magma_CSR) {
magma_z_mfree( &A_h1, queue );
magma_z_mfree( &A_h2, queue );
}
magma_z_mfree( &B, queue );
magma_z_mfree( &C, queue );
magma_z_vfree( &diag, queue );
magma_z_vfree( &c_t, queue );
magma_z_vfree( &b_h, queue );
return MAGMA_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_zbaiter
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_z_solver_par solver_par;
magma_z_preconditioner precond_par;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_z_sparse_matrix A;
magma_z_vector x, b;
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
int i;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--maxiter", argv[i]) == 0 ){
solver_par.maxiter = atoi( argv[++i] );
}else if ( strcmp("--mscale", argv[i]) == 0 ) {
scale = atoi( argv[++i] );
switch( scale ) {
case 0: scaling = Magma_NOSCALE; break;
case 1: scaling = Magma_UNITDIAG; break;
case 2: scaling = Magma_UNITROW; break;
}
}else
break;
}
printf( "\n# usage: ./run_zbaiter"
" [ "
" --mscale %d (0=no, 1=unitdiag, 2=unitrownrm)"
" --maxiter %d ]"
" matrices \n\n",
(int) scale,
(int) solver_par.maxiter);
magma_zsolverinfo_init( &solver_par, &precond_par );
while( i < argc ){
magma_z_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 initial guess
magma_zmscale( &A, scaling );
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zbaiter( A, b, &x, &solver_par );
magma_zsolverinfo( &solver_par, &precond_par );
magma_z_mfree(&A);
magma_z_vfree(&x);
magma_z_vfree(&b);
i++;
}
magma_zsolverinfo_free( &solver_par, &precond_par );
TESTING_FINALIZE();
return 0;
}
magma_int_t
magma_zpcg( magma_z_sparse_matrix A, magma_z_vector b, magma_z_vector *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par ){
// prepare solver feedback
solver_par->solver = Magma_PCG;
solver_par->numiter = 0;
solver_par->info = 0;
// local variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_z_vector r, rt, p, q, h;
magma_z_vinit( &r, Magma_DEV, dofs, c_zero );
magma_z_vinit( &rt, Magma_DEV, dofs, c_zero );
magma_z_vinit( &p, Magma_DEV, dofs, c_zero );
magma_z_vinit( &q, Magma_DEV, dofs, c_zero );
magma_z_vinit( &h, Magma_DEV, dofs, c_zero );
// solver variables
magmaDoubleComplex alpha, beta;
double nom, nom0, r0, gammaold, gammanew, den, res;
// solver setup
magma_zscal( dofs, c_zero, x->val, 1) ; // x = 0
magma_zcopy( dofs, b.val, 1, r.val, 1 ); // r = b
// preconditioner
magma_z_applyprecond_left( A, r, &rt, precond_par );
magma_z_applyprecond_right( A, rt, &h, precond_par );
magma_zcopy( dofs, h.val, 1, p.val, 1 ); // p = h
nom = MAGMA_Z_REAL( magma_zdotc(dofs, r.val, 1, h.val, 1) );
nom0 = magma_dznrm2( dofs, r.val, 1 );
magma_z_spmv( c_one, A, p, c_zero, q ); // q = A p
den = MAGMA_Z_REAL( magma_zdotc(dofs, p.val, 1, q.val, 1) );// den = p dot q
solver_par->init_res = nom0;
if ( (r0 = nom * solver_par->epsilon) < ATOLERANCE )
r0 = ATOLERANCE;
if ( nom < r0 )
return MAGMA_SUCCESS;
// check positive definite
if (den <= 0.0) {
printf("Operator A is not postive definite. (Ar,r) = %f\n", den);
return -100;
}
//Chronometry
real_Double_t tempo1, tempo2;
magma_device_sync(); tempo1=magma_wtime();
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++ ){
// preconditioner
magma_z_applyprecond_left( A, r, &rt, precond_par );
magma_z_applyprecond_right( A, rt, &h, precond_par );
gammanew = MAGMA_Z_REAL( magma_zdotc(dofs, r.val, 1, h.val, 1) );
// gn = < r,h>
if( solver_par->numiter==1 ){
magma_zcopy( dofs, h.val, 1, p.val, 1 ); // p = h
}else{
beta = MAGMA_Z_MAKE(gammanew/gammaold, 0.); // beta = gn/go
magma_zscal(dofs, beta, p.val, 1); // p = beta*p
magma_zaxpy(dofs, c_one, h.val, 1, p.val, 1); // p = p + h
}
magma_z_spmv( c_one, A, p, c_zero, q ); // q = A p
den = MAGMA_Z_REAL(magma_zdotc(dofs, p.val, 1, q.val, 1));
// den = p dot q
alpha = MAGMA_Z_MAKE(gammanew/den, 0.);
magma_zaxpy(dofs, alpha, p.val, 1, x->val, 1); // x = x + alpha p
magma_zaxpy(dofs, -alpha, q.val, 1, r.val, 1); // r = r - alpha q
gammaold = gammanew;
res = magma_dznrm2( dofs, r.val, 1 );
if( solver_par->verbose > 0 ){
magma_device_sync(); tempo2=magma_wtime();
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;
}
}
magma_device_sync(); tempo2=magma_wtime();
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
magma_zresidual( A, b, *x, &residual );
solver_par->iter_res = res;
solver_par->final_res = residual;
if( solver_par->numiter < solver_par->maxiter){
solver_par->info = 0;
}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;
}
}
solver_par->info = -2;
}
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;
}
}
solver_par->info = -1;
}
magma_z_vfree(&r);
magma_z_vfree(&rt);
magma_z_vfree(&p);
magma_z_vfree(&q);
magma_z_vfree(&h);
return MAGMA_SUCCESS;
} /* magma_zcg */
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_zgmres
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_z_solver_par solver_par;
magma_z_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.restart = 30;
solver_par.num_eigenvalues = 0;
solver_par.ortho = Magma_CGS;
solver_par.verbose = 0;
int format = 0;
int ortho = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_z_sparse_matrix A, B, B_d;
magma_z_vector x, b;
B.blocksize = 8;
B.alignment = 8;
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
B.storage_type = Magma_CSR;
int i;
for( i = 1; i < argc; ++i ) {
if ( strcmp("--format", argv[i]) == 0 ) {
format = atoi( argv[++i] );
switch( format ) {
case 0: B.storage_type = Magma_CSR; break;
case 1: B.storage_type = Magma_ELL; break;
case 2: B.storage_type = Magma_ELLRT; break;
case 3: B.storage_type = Magma_SELLP; break;
}
}else if ( strcmp("--mscale", argv[i]) == 0 ) {
scale = atoi( argv[++i] );
switch( scale ) {
case 0: scaling = Magma_NOSCALE; break;
case 1: scaling = Magma_UNITDIAG; break;
case 2: scaling = Magma_UNITROW; break;
}
}else if ( strcmp("--blocksize", argv[i]) == 0 ) {
B.blocksize = atoi( argv[++i] );
}else if ( strcmp("--alignment", argv[i]) == 0 ) {
B.alignment = atoi( argv[++i] );
}else if ( strcmp("--verbose", argv[i]) == 0 ) {
solver_par.verbose = atoi( argv[++i] );
} else if ( strcmp("--ortho", argv[i]) == 0 ) {
ortho = atoi( argv[++i] );
switch( ortho ) {
case 0: solver_par.ortho = Magma_CGS; break;
case 1: solver_par.ortho = Magma_MGS; break;
case 2: solver_par.ortho = Magma_FUSED_CGS; break;
}
} else if ( strcmp("--restart", argv[i]) == 0 ) {
solver_par.restart = atoi( argv[++i] );
} else if ( strcmp("--maxiter", argv[i]) == 0 ) {
solver_par.maxiter = atoi( argv[++i] );
} else if ( strcmp("--tol", argv[i]) == 0 ) {
sscanf( argv[++i], "%lf", &solver_par.epsilon );
} else
break;
}
printf( "\n# usage: ./run_zgmres"
" [ --format %d (0=CSR, 1=ELL 2=ELLRT, 3=SELLP)"
" [ --blocksize %d --alignment %d ]"
" --mscale %d (0=no, 1=unitdiag, 2=unitrownrm)"
" --verbose %d (0=summary, k=details every k iterations)"
" --restart %d --maxiter %d --tol %.2e"
" --ortho %d (0=CGS, 1=MGS, 2=FUSED_CGS) ]"
" matrices \n\n", format, (int) B.blocksize, (int) B.alignment,
(int) scale,
(int) solver_par.verbose,
(int) solver_par.restart, (int) solver_par.maxiter,
solver_par.epsilon, ortho );
magma_zsolverinfo_init( &solver_par, &precond_par );
while( i < argc ){
magma_z_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_zmscale( &A, scaling );
magma_z_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zgmres( B_d, b, &x, &solver_par );
magma_zsolverinfo( &solver_par, &precond_par );
magma_z_mfree(&B_d);
magma_z_mfree(&B);
magma_z_mfree(&A);
magma_z_vfree(&x);
magma_z_vfree(&b);
i++;
}
magma_zsolverinfo_free( &solver_par, &precond_par );
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Debugging file
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_z_solver_par solver_par;
magma_z_preconditioner precond_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.restart = 30;
solver_par.num_eigenvalues = 0;
solver_par.ortho = Magma_CGS;
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_z_sparse_matrix A, B, B_d;
magma_z_vector x, b;
// generate matrix of desired structure and size
magma_int_t n=100; // size is n*n
magma_int_t nn = n*n;
magma_int_t offdiags = 2;
magma_index_t *diag_offset;
magmaDoubleComplex *diag_vals;
magma_zmalloc_cpu( &diag_vals, offdiags+1 );
magma_index_malloc_cpu( &diag_offset, offdiags+1 );
diag_offset[0] = 0;
diag_offset[1] = 1;
diag_offset[2] = n;
diag_vals[0] = MAGMA_Z_MAKE( 4.1, 0.0 );
diag_vals[1] = MAGMA_Z_MAKE( -1.0, 0.0 );
diag_vals[2] = MAGMA_Z_MAKE( -1.0, 0.0 );
magma_zmgenerator( nn, offdiags, diag_offset, diag_vals, &A );
// convert marix into desired format
B.storage_type = Magma_SELLC;
B.blocksize = 8;
B.alignment = 8;
// scale matrix
magma_zmscale( &A, Magma_UNITDIAG );
magma_z_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_z_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// test CG ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zcg_res( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PCG IC ################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ICC;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpcg( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test BICGSTAB ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zbicgstab( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PBICGSTAB Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
/*
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PBICGSTAB ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);printf("here\n");
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpbicgstab( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test GMRES ####################################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// solver
magma_zgmres( B_d, b, &x, &solver_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
// test PGMRES Jacobi ############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_JACOBI;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);*/
// test PGMRES ILU ###############################
// vectors and initial guess
magma_z_vinit( &b, Magma_DEV, A.num_cols, one );
magma_z_vinit( &x, Magma_DEV, A.num_cols, one );
magma_z_spmv( one, B_d, x, zero, b ); // b = A x
magma_z_vfree(&x);
magma_z_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_zsolverinfo_init( &solver_par, &precond_par );
// Preconditioner
precond_par.solver = Magma_ILU;
magma_z_precondsetup( B_d, b, &precond_par );
// solver
magma_zpgmres( B_d, b, &x, &solver_par, &precond_par );
// solverinfo
magma_zsolverinfo( &solver_par, &precond_par );
if( solver_par.numiter > 150 ){
printf("error: test not passed!\n"); exit(-1);
}
magma_zsolverinfo_free( &solver_par, &precond_par );
magma_z_vfree(&x);
magma_z_vfree(&b);
printf("all tests passed.\n");
magma_z_mfree(&B_d);
magma_z_mfree(&B);
magma_z_mfree(&A);
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- 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_z_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;
magmaDoubleComplex c_one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex c_zero = MAGMA_Z_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_zspmv"
" [ --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_zm_5stencil( laplace_size, &hA, queue );
} else { // file-matrix test
magma_z_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_z_vector hx, hy, dx, dy, hrefvec, hcheck;
// init CPU vectors
magma_z_vinit( &hx, Magma_CPU, hA.num_rows, c_zero, queue );
magma_z_vinit( &hy, Magma_CPU, hA.num_rows, c_zero, queue );
// init DEV vectors
magma_z_vinit( &dx, Magma_DEV, hA.num_rows, c_one, queue );
magma_z_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_zcsrmv( "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_z_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_z_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_z_mfree(&dA, queue );
magma_z_vtransfer( dy, &hrefvec , Magma_DEV, Magma_CPU, queue );
// convert to ELL and copy to GPU
magma_z_mconvert( hA, &hA_ELL, Magma_CSR, Magma_ELL, queue );
magma_z_mtransfer( hA_ELL, &dA_ELL, Magma_CPU, Magma_DEV, queue );
magma_z_mfree(&hA_ELL, queue );
magma_z_vfree( &dy, queue );
magma_z_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_z_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_z_mfree(&dA_ELL, queue );
magma_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_REAL(hrefvec.val[k]);
if ( res < .000001 )
printf("# tester spmv ELL: ok\n");
else
printf("# tester spmv ELL: failed\n");
magma_z_vfree( &hcheck, queue );
// convert to SELLP and copy to GPU
magma_z_mconvert( hA, &hA_SELLP, Magma_CSR, Magma_SELLP, queue );
magma_z_mtransfer( hA_SELLP, &dA_SELLP, Magma_CPU, Magma_DEV, queue );
magma_z_mfree(&hA_SELLP, queue );
magma_z_vfree( &dy, queue );
magma_z_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_z_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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue );
magma_z_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);
magmaDoubleComplex alpha = c_one;
magmaDoubleComplex beta = c_zero;
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
// copy matrix to GPU
magma_z_mtransfer( hA, &dA, Magma_CPU, Magma_DEV, queue );
start = magma_sync_wtime( queue );
for (j=0; j<10; j++)
cusparseStatus =
cusparseZcsrmv(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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue );
magma_z_vfree( &dy, queue );
magma_z_vinit( &dy, Magma_DEV, hA.num_rows, c_zero, queue );
cusparseZcsr2hyb(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 =
cusparseZhybmv( 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_z_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_Z_REAL(hcheck.val[k]) - MAGMA_Z_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_z_vfree( &hcheck, queue );
cusparseDestroyMatDescr( descrA );
cusparseDestroyHybMat( hybA );
cusparseDestroy( cusparseHandle );
magma_z_mfree(&dA, queue );
printf("\n\n");
// free CPU memory
magma_z_mfree(&hA, queue );
magma_z_vfree(&hx, queue );
magma_z_vfree(&hy, queue );
magma_z_vfree(&hrefvec, queue );
// free GPU memory
magma_z_vfree(&dx, queue );
magma_z_vfree(&dy, queue );
i++;
}
magma_queue_destroy( queue );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_zcg(
magma_z_sparse_matrix A, magma_z_vector b, magma_z_vector *x,
magma_z_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
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_z_vector r, p, q;
magma_z_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_z_vinit( &p, Magma_DEV, dofs, c_zero, queue );
magma_z_vinit( &q, Magma_DEV, dofs, c_zero, queue );
// solver variables
magmaDoubleComplex alpha, beta;
double nom, nom0, r0, betanom, betanomsq, den;
// solver setup
magma_zscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_zcopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
magma_zcopy( dofs, b.dval, 1, p.dval, 1 ); // p = b
nom0 = betanom = magma_dznrm2( dofs, r.dval, 1 );
nom = nom0 * nom0; // nom = r' * r
magma_z_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_Z_REAL( magma_zdotc(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_Z_MAKE(nom/den, 0.);
magma_zaxpy(dofs, alpha, p.dval, 1, x->dval, 1); // x = x + alpha p
magma_zaxpy(dofs, -alpha, q.dval, 1, r.dval, 1); // r = r - alpha q
betanom = magma_dznrm2(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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < r0 ) {
break;
}
beta = MAGMA_Z_MAKE(betanomsq/nom, 0.); // beta = betanoms/nom
magma_zscal(dofs, beta, p.dval, 1); // p = beta*p
magma_zaxpy(dofs, c_one, r.dval, 1, p.dval, 1); // p = p + r
magma_z_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_Z_REAL(magma_zdotc(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;
double residual;
magma_zresidual( A, b, *x, &residual, queue );
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_z_vfree(&r, queue );
magma_z_vfree(&p, queue );
magma_z_vfree(&q, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_zcg */
/* ////////////////////////////////////////////////////////////////////////////
-- testing zdot
*/
int main( int argc, char** argv)
{
TESTING_INIT();
printf("#================================================================================================================================================\n");
printf("\n");
printf(" | runtime | GFLOPS\n");
printf("#n num_vecs | CUDOT CUGEMV MAGMAGEMV MDOT MDGM | CUDOT CUGEMV MAGMAGEMV MDOT MDGM \n");
printf("#------------------------------------------------------------------------------------------------------------------------------------------------\n");
printf("\n");
for( magma_int_t num_vecs=5; num_vecs<6; num_vecs+=1 ){
for( magma_int_t n=10000; n<100000001; n=n+10000 ){
magma_z_sparse_matrix A, B, C, D, E, F, G, H, I, J, K, Z;
magma_z_vector a,b,c,x, y, z, skp;
int iters = 10;
double computations = (2.* n * iters * num_vecs);
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magmaDoubleComplex alpha;
#define ENABLE_TIMER
#ifdef ENABLE_TIMER
double mdot1, mdot2, mdgm1, mdgm2, magmagemv1, magmagemv2, cugemv1, cugemv2, cudot1, cudot2;
double mdot_time, mdgm_time, magmagemv_time, cugemv_time, cudot_time;
#endif
magma_z_vinit( &a, Magma_DEV, n*num_vecs, one );
magma_z_vinit( &b, Magma_DEV, num_vecs, one );
int min_ten = min(num_vecs, 15);
magma_z_vinit( &x, Magma_DEV, min_ten*n, one );
magma_z_vinit( &y, Magma_DEV, min_ten*n, one );
magma_z_vinit( &skp, Magma_DEV, num_vecs, zero );
// warm up
magma_zgemvmdot( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
// CUDOT
#ifdef ENABLE_TIMER
magma_device_sync(); cudot1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
for( int l=0; l<num_vecs; l++)
alpha = magma_zdotc(n, a.val, 1, b.val, 1);
}
#ifdef ENABLE_TIMER
magma_device_sync(); cudot2=magma_wtime();
cudot_time=cudot2-cudot1;
#endif
// CUGeMV
#ifdef ENABLE_TIMER
magma_device_sync(); cugemv1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magma_zgemv(MagmaTrans, n, num_vecs, one, a.val, n, b.val, 1, zero, skp.val, 1);
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); cugemv2=magma_wtime();
cugemv_time=cugemv2-cugemv1;
#endif
// MAGMAGeMV
#ifdef ENABLE_TIMER
magma_device_sync(); magmagemv1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magmablas_zgemv(MagmaTrans, n, num_vecs, one, a.val, n, b.val, 1, zero, skp.val, 1);
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); magmagemv2=magma_wtime();
magmagemv_time=magmagemv2-magmagemv1;
#endif
// MDOT
#ifdef ENABLE_TIMER
magma_device_sync(); mdot1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
//magma_zmdotc( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 2, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 2, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 1, a.val, b.val, x.val, y.val, skp.val );
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); mdot2=magma_wtime();
mdot_time=mdot2-mdot1;
#endif
// MDGM
#ifdef ENABLE_TIMER
magma_device_sync(); mdgm1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magma_zgemvmdot( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); mdgm2=magma_wtime();
mdgm_time=mdgm2-mdgm1;
#endif
//magma_zprint_gpu(num_vecs,1,skp.val,num_vecs);
//Chronometry
#ifdef ENABLE_TIMER
printf("%d %d %e %e %e %e %e %e %e %e %e %e\n",
n, num_vecs,
cudot_time/iters,
(cugemv_time)/iters,
(magmagemv_time)/iters,
(mdot_time)/iters,
(mdgm_time)/iters,
(double)(computations)/(cudot_time*(1.e+09)),
(double)(computations)/(cugemv_time*(1.e+09)),
(double)(computations)/(magmagemv_time*(1.e+09)),
(double)(computations)/(mdot_time*(1.e+09)),
(double)(computations)/(mdgm_time*(1.e+09)) );
#endif
magma_z_vfree(&a);
magma_z_vfree(&b);
magma_z_vfree(&x);
magma_z_vfree(&y);
magma_z_vfree(&skp);
}
// }
printf("#================================================================================================================================================\n");
printf("\n");
printf("\n");
}
TESTING_FINALIZE();
return 0;
}
