magma_int_t
magma_sresidual( magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector x,
float *res ){
// some useful variables
float zero = MAGMA_S_ZERO, one = MAGMA_S_ONE,
mone = MAGMA_S_NEG_ONE;
magma_int_t dofs = A.num_rows;
magma_s_vector r;
magma_s_vinit( &r, Magma_DEV, A.num_rows, zero );
magma_s_spmv( one, A, x, zero, r ); // r = A x
magma_saxpy(dofs, mone, b.val, 1, r.val, 1); // r = r - b
*res = magma_snrm2(dofs, r.val, 1); // res = ||r||
// /magma_snrm2(dofs, b.val, 1); /||b||
//printf( "relative residual: %e\n", *res );
magma_s_vfree(&r);
return MAGMA_SUCCESS;
}
magma_int_t
magma_spastixsetup( magma_s_sparse_matrix A, magma_s_vector b,
magma_s_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_s_sparse_matrix A_h1, B;
magma_s_vector diag, c_t, b_h;
magma_s_vinit( &c_t, Magma_CPU, A.num_rows, MAGMA_S_ZERO );
magma_s_vinit( &diag, Magma_CPU, A.num_rows, MAGMA_S_ZERO );
magma_s_vtransfer( b, &b_h, A.memory_location, Magma_CPU);
if( A.storage_type != Magma_CSR ){
magma_s_mtransfer( A, &A_h1, A.memory_location, Magma_CPU);
magma_s_mconvert( A_h1, &B, A_h1.storage_type, Magma_CSR);
}
else{
magma_s_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 = (float*) 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_s_mfree( &A_h1 );
}
magma_s_vfree( &b_h);
magma_s_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_sapplypastix( magma_s_vector b, magma_s_vector *x,
magma_s_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_s_vector b_h;
magma_s_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_ssetvector( ncol,
// b_h.val, 1, x->val, 1 );
magma_s_vfree( x );
magma_s_vtransfer( b_h, x, Magma_CPU, b.memory_location);
magma_s_vfree( &b_h);
#else
printf( "error: only real supported yet.\n");
#endif
#else
printf( "error: pastix not available.\n");
#endif
return MAGMA_SUCCESS;
}
magma_int_t
magma_spgmres( magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector *x,
magma_s_solver_par *solver_par,
magma_s_preconditioner *precond_par ){
// prepare solver feedback
solver_par->solver = Magma_PGMRES;
solver_par->numiter = 0;
solver_par->info = 0;
// local variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_ONE,
c_mone = MAGMA_S_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);
float *H, *HH, *y, *h1;
magma_smalloc_pinned( &H, (ldh+1)*ldh );
magma_smalloc_pinned( &y, ldh );
magma_smalloc_pinned( &HH, ldh*ldh );
magma_smalloc_pinned( &h1, ldh );
// GPU workspace
magma_s_vector r, q, q_t, z, z_t, t;
magma_s_vinit( &t, Magma_DEV, dofs, c_zero );
magma_s_vinit( &r, Magma_DEV, dofs, c_zero );
magma_s_vinit( &q, Magma_DEV, dofs*(ldh+1), c_zero );
magma_s_vinit( &z, Magma_DEV, dofs*(ldh+1), c_zero );
magma_s_vinit( &z_t, Magma_DEV, dofs, c_zero );
q_t.memory_location = Magma_DEV;
q_t.val = NULL;
q_t.num_rows = q_t.nnz = dofs;
float *dy, *dH = NULL;
if (MAGMA_SUCCESS != magma_smalloc( &dy, ldh ))
return MAGMA_ERR_DEVICE_ALLOC;
if (MAGMA_SUCCESS != magma_smalloc( &dH, (ldh+1)*ldh ))
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_sscal( dofs, c_zero, x->val, 1 ); // x = 0
magma_scopy( dofs, b.val, 1, r.val, 1 ); // r = b
nom0 = betanom = magma_snrm2( dofs, r.val, 1 ); // nom0= || r||
nom = nom0 * nom0;
solver_par->init_res = nom0;
H(1,0) = MAGMA_S_MAKE( nom0, 0. );
magma_ssetvector(1, &H(1,0), 1, &dH(1,0), 1);
if ( (r0 = nom0 * RTOLERANCE ) < ATOLERANCE )
r0 = solver_par->epsilon;
if ( nom < r0 )
return MAGMA_SUCCESS;
//Chronometry
real_Double_t tempo1, tempo2;
magma_device_sync(); tempo1=magma_wtime();
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_scopy(dofs, r.val, 1, q(k-1), 1); // q[0] = 1.0/||r||
magma_sscal(dofs, 1./H(k,k-1), q(k-1), 1); // (to be fused)
q_t.val = q(k-1);
magmablasSetKernelStream(stream[0]);
// preconditioner
// z[k] = M^(-1) q(k)
magma_s_applyprecond_left( A, q_t, &t, precond_par );
magma_s_applyprecond_right( A, t, &z_t, precond_par );
magma_scopy(dofs, z_t.val, 1, z(k-1), 1);
// r = A q[k]
magma_s_spmv( c_one, A, z_t, c_zero, r );
// if (solver_par->ortho == Magma_MGS ) {
// modified Gram-Schmidt
for (i=1; i<=k; i++) {
H(i,k) =magma_sdot(dofs, q(i-1), 1, r.val, 1);
// H(i,k) = q[i] . r
magma_saxpy(dofs,-H(i,k), q(i-1), 1, r.val, 1);
// r = r - H(i,k) q[i]
}
H(k+1,k) = MAGMA_S_MAKE( magma_snrm2(dofs, r.val, 1), 0. ); // H(k+1,k) = ||r||
/*}else if (solver_par->ortho == Magma_FUSED_CGS ) {
// fusing sgemv with snrm2 in classical Gram-Schmidt
magmablasSetKernelStream(stream[0]);
magma_scopy(dofs, r.val, 1, q(k), 1);
// dH(1:k+1,k) = q[0:k] . r
magmablas_sgemv(MagmaTrans, dofs, k+1, c_one, q(0),
dofs, r.val, 1, c_zero, &dH(1,k), 1);
// r = r - q[0:k-1] dH(1:k,k)
magmablas_sgemv(MagmaNoTrans, dofs, k, c_mone, q(0),
dofs, &dH(1,k), 1, c_one, r.val, 1);
// 1) dH(k+1,k) = sqrt( dH(k+1,k) - dH(1:k,k) )
magma_scopyscale( dofs, k, r.val, 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_sgetvector_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_sgemv(MagmaTrans, dofs, k, c_one, q(0),
dofs, r.val, 1, c_zero, &dH(1,k), 1);
// dH(1:k,k) = q[0:k-1] . r
#ifndef SNRM2SCALE
// 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_sgetvector_async(k, &dH(1,k), 1, &H(1,k),
1, stream[1]);
#endif
// r = r - q[0:k-1] dH(1:k,k)
magmablas_sgemv(MagmaNoTrans, dofs, k, c_mone, q(0),
dofs, &dH(1,k), 1, c_one, r.val, 1);
#ifdef SNRM2SCALE
magma_scopy(dofs, r.val, 1, q(k), 1);
// q[k] = r / H(k,k-1)
magma_snrm2scale(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_sgetvector_async(k+1, &dH(1,k), 1, &H(1,k), 1, stream[1]);
#else
H(k+1,k) = MAGMA_S_MAKE( magma_snrm2(dofs, r.val, 1), 0. );
// H(k+1,k) = sqrt(r . r)
if( k<solver_par->restart ){
magmablasSetKernelStream(stream[0]);
magma_scopy(dofs, r.val, 1, q(k), 1);
// q[k] = 1.0/H[k][k-1] r
magma_sscal(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_sdot( 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_S_REAL(H(k+1,k)));
}/* Minimization done */
// compute solution approximation
magma_ssetmatrix(m, 1, y+1, m, dy, m );
magma_sgemv(MagmaNoTrans, dofs, m, c_one, z(0), dofs, dy, 1,
c_one, x->val, 1);
// compute residual
magma_s_spmv( c_mone, A, *x, c_zero, r ); // r = - A * x
magma_saxpy(dofs, c_one, b.val, 1, r.val, 1); // r = r + b
H(1,0) = MAGMA_S_MAKE( magma_snrm2(dofs, r.val, 1), 0. );
// RNorm = H[1][0] = || r ||
RNorm = MAGMA_S_REAL( H(1,0) );
betanom = fabs(RNorm);
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < r0 ) {
break;
}
}
magma_device_sync(); tempo2=magma_wtime();
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_sresidual( A, b, *x, &residual );
solver_par->iter_res = betanom;
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) betanom;
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = -1;
}
// 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_s_vfree(&t);
magma_s_vfree(&r);
magma_s_vfree(&q);
magma_s_vfree(&z);
magma_s_vfree(&z_t);
// free GPU streams and events
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_event_destroy( event[0] );
magmablasSetKernelStream(NULL);
return MAGMA_SUCCESS;
} /* magma_spgmres */
extern "C" magma_int_t
magma_scg_merge(
magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector *x,
magma_s_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_CGMERGE;
solver_par->numiter = 0;
solver_par->info = MAGMA_SUCCESS;
// some useful variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_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] );
// GPU workspace
magma_s_vector r, d, z;
magma_s_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_s_vinit( &d, Magma_DEV, dofs, c_zero, queue );
magma_s_vinit( &z, Magma_DEV, dofs, c_zero, queue );
float *d1, *d2, *skp;
d1 = NULL;
d2 = NULL;
skp = NULL;
magma_int_t stat_dev = 0, stat_cpu = 0;
stat_dev += magma_smalloc( &d1, dofs*(1) );
stat_dev += magma_smalloc( &d2, dofs*(1) );
// array for the parameters
stat_dev += magma_smalloc( &skp, 6 );
// skp = [alpha|beta|gamma|rho|tmp1|tmp2]
if( stat_dev != 0 ){
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
printf("error: memory allocation.\n");
return MAGMA_ERR_DEVICE_ALLOC;
}
// solver variables
float alpha, beta, gamma, rho, tmp1, *skp_h;
float nom, nom0, r0, betanom, den;
// solver setup
magma_sscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_scopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
magma_scopy( dofs, b.dval, 1, d.dval, 1 ); // d = b
nom0 = betanom = magma_snrm2( dofs, r.dval, 1 );
nom = nom0 * nom0; // nom = r' * r
magma_s_spmv( c_one, A, d, c_zero, z, queue ); // z = A d
den = MAGMA_S_REAL( magma_sdot(dofs, d.dval, 1, z.dval, 1) ); // den = d'* z
solver_par->init_res = nom0;
// array on host for the parameters
stat_cpu += magma_smalloc_cpu( &skp_h, 6 );
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;
}
alpha = rho = gamma = tmp1 = c_one;
beta = magma_sdot(dofs, r.dval, 1, r.dval, 1);
skp_h[0]=alpha;
skp_h[1]=beta;
skp_h[2]=gamma;
skp_h[3]=rho;
skp_h[4]=tmp1;
skp_h[5]=MAGMA_S_MAKE(nom, 0.0);
magma_ssetvector( 6, skp_h, 1, skp, 1 );
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++ ) {
magmablasSetKernelStream(stream[0]);
// computes SpMV and dot product
magma_scgmerge_spmv1( A, d1, d2, d.dval, z.dval, skp, queue );
// updates x, r, computes scalars and updates d
magma_scgmerge_xrbeta( dofs, d1, d2, x->dval, r.dval, d.dval, z.dval, skp, queue );
// check stopping criterion (asynchronous copy)
magma_sgetvector_async( 1 , skp+1, 1,
skp_h+1, 1, stream[1] );
betanom = sqrt(MAGMA_S_REAL(skp_h[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) 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_sresidual( 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_s_vfree(&r, queue );
magma_s_vfree(&z, queue );
magma_s_vfree(&d, queue );
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
magma_free_cpu( skp_h );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_scg_merge */
magma_int_t
magma_sbicgstab( magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector *x,
magma_s_solver_par *solver_par ){
// prepare solver feedback
solver_par->solver = Magma_BICGSTAB;
solver_par->numiter = 0;
solver_par->info = 0;
// some useful variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_ONE,
c_mone = MAGMA_S_NEG_ONE;
magma_int_t dofs = A.num_rows;
// workspace
magma_s_vector r,rr,p,v,s,t;
magma_s_vinit( &r, Magma_DEV, dofs, c_zero );
magma_s_vinit( &rr, Magma_DEV, dofs, c_zero );
magma_s_vinit( &p, Magma_DEV, dofs, c_zero );
magma_s_vinit( &v, Magma_DEV, dofs, c_zero );
magma_s_vinit( &s, Magma_DEV, dofs, c_zero );
magma_s_vinit( &t, Magma_DEV, dofs, c_zero );
// solver variables
float alpha, beta, omega, rho_old, rho_new;
float nom, betanom, nom0, r0, den, res;
// solver setup
magma_sscal( dofs, c_zero, x->val, 1) ; // x = 0
magma_scopy( dofs, b.val, 1, r.val, 1 ); // r = b
magma_scopy( dofs, b.val, 1, rr.val, 1 ); // rr = b
nom0 = betanom = magma_snrm2( dofs, r.val, 1 ); // nom = || r ||
nom = nom0*nom0;
rho_old = omega = alpha = MAGMA_S_MAKE( 1.0, 0. );
solver_par->init_res = nom0;
magma_s_spmv( c_one, A, r, c_zero, v ); // z = A r
den = MAGMA_S_REAL( magma_sdot(dofs, v.val, 1, r.val, 1) ); // den = z' * r
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] = 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_sdot( dofs, rr.val, 1, r.val, 1 ); // rho=<rr,r>
beta = rho_new/rho_old * alpha/omega; // beta=rho/rho_old *alpha/omega
magma_sscal( dofs, beta, p.val, 1 ); // p = beta*p
magma_saxpy( dofs, c_mone * omega * beta, v.val, 1 , p.val, 1 );
// p = p-omega*beta*v
magma_saxpy( dofs, c_one, r.val, 1, p.val, 1 ); // p = p+r
magma_s_spmv( c_one, A, p, c_zero, v ); // v = Ap
alpha = rho_new / magma_sdot( dofs, rr.val, 1, v.val, 1 );
magma_scopy( dofs, r.val, 1 , s.val, 1 ); // s=r
magma_saxpy( dofs, c_mone * alpha, v.val, 1 , s.val, 1 ); // s=s-alpha*v
magma_s_spmv( c_one, A, s, c_zero, t ); // t=As
omega = magma_sdot( dofs, t.val, 1, s.val, 1 ) // omega = <s,t>/<t,t>
/ magma_sdot( dofs, t.val, 1, t.val, 1 );
magma_saxpy( dofs, alpha, p.val, 1 , x->val, 1 ); // x=x+alpha*p
magma_saxpy( dofs, omega, s.val, 1 , x->val, 1 ); // x=x+omega*s
magma_scopy( dofs, s.val, 1 , r.val, 1 ); // r=s
magma_saxpy( dofs, c_mone * omega, t.val, 1 , r.val, 1 ); // r=r-omega*t
res = betanom = magma_snrm2( dofs, r.val, 1 );
nom = betanom*betanom;
rho_old = rho_new; // rho_old=rho
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;
float residual;
magma_sresidual( 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) betanom;
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = -1;
}
magma_s_vfree(&r);
magma_s_vfree(&rr);
magma_s_vfree(&p);
magma_s_vfree(&v);
magma_s_vfree(&s);
magma_s_vfree(&t);
return MAGMA_SUCCESS;
} /* magma_sbicgstab */
magma_int_t
magma_sbicgstab_merge( magma_s_sparse_matrix A, magma_s_vector b,
magma_s_vector *x, magma_s_solver_par *solver_par ){
// prepare solver feedback
solver_par->solver = Magma_BICGSTABMERGE;
solver_par->numiter = 0;
solver_par->info = 0;
// some useful variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_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_s_vector q, r,rr,p,v,s,t;
float *d1, *d2, *skp;
magma_smalloc( &d1, dofs*(2) );
magma_smalloc( &d2, dofs*(2) );
// array for the parameters
magma_smalloc( &skp, 8 );
// skp = [alpha|beta|omega|rho_old|rho|nom|tmp1|tmp2]
magma_s_vinit( &q, Magma_DEV, dofs*6, c_zero );
// q = rr|r|p|v|s|t
rr.memory_location = Magma_DEV; rr.val = NULL; rr.num_rows = rr.nnz = dofs;
r.memory_location = Magma_DEV; r.val = NULL; r.num_rows = r.nnz = dofs;
p.memory_location = Magma_DEV; p.val = NULL; p.num_rows = p.nnz = dofs;
v.memory_location = Magma_DEV; v.val = NULL; v.num_rows = v.nnz = dofs;
s.memory_location = Magma_DEV; s.val = NULL; s.num_rows = s.nnz = dofs;
t.memory_location = Magma_DEV; t.val = NULL; t.num_rows = t.nnz = dofs;
rr.val = q(0);
r.val = q(1);
p.val = q(2);
v.val = q(3);
s.val = q(4);
t.val = q(5);
// solver variables
float alpha, beta, omega, rho_old, rho_new, *skp_h;
float nom, nom0, betanom, r0, den;
// solver setup
magma_sscal( dofs, c_zero, x->val, 1) ; // x = 0
magma_scopy( dofs, b.val, 1, q(0), 1 ); // rr = b
magma_scopy( dofs, b.val, 1, q(1), 1 ); // r = b
rho_new = magma_sdot( dofs, r.val, 1, r.val, 1 ); // rho=<rr,r>
nom = MAGMA_S_REAL(magma_sdot( dofs, r.val, 1, r.val, 1 ));
nom0 = betanom = sqrt(nom); // nom = || r ||
rho_old = omega = alpha = MAGMA_S_MAKE( 1.0, 0. );
beta = rho_new;
solver_par->init_res = nom0;
// array on host for the parameters
magma_smalloc_cpu( &skp_h, 8 );
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_S_MAKE(nom, 0.0);
magma_ssetvector( 8, skp_h, 1, skp, 1 );
magma_s_spmv( c_one, A, r, c_zero, v ); // z = A r
den = MAGMA_S_REAL( magma_sdot(dofs, v.val, 1, r.val, 1) );// den = z dot r
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] = 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_sbicgmerge1( dofs, skp, v.val, r.val, p.val );
magma_s_spmv( c_one, A, p, c_zero, v ); // v = Ap
magma_smdotc( dofs, 1, q.val, v.val, d1, d2, skp );
magma_sbicgmerge4( 1, skp );
magma_sbicgmerge2( dofs, skp, r.val, v.val, s.val ); // s=r-alpha*v
magma_s_spmv( c_one, A, s, c_zero, t ); // t=As
magma_smdotc( dofs, 2, q.val+4*dofs, t.val, d1, d2, skp+6 );
magma_sbicgmerge4( 2, skp );
magma_sbicgmerge3( dofs, skp, p.val, s.val, // x=x+alpha*p+omega*s
t.val, x->val, r.val ); // r=s-omega*t
magma_smdotc( dofs, 2, q.val, r.val, d1, d2, skp+4);
magma_sbicgmerge4( 3, skp );
// check stopping criterion (asynchronous copy)
magma_sgetvector_async( 1 , skp+5, 1,
skp_h+5, 1, stream[1] );
betanom = sqrt(MAGMA_S_REAL(skp_h[5]));
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < r0 ) {
break;
}
}
magma_device_sync(); tempo2=magma_wtime();
solver_par->runtime = (real_Double_t) tempo2-tempo1;
float residual;
magma_sresidual( A, b, *x, &residual );
solver_par->iter_res = betanom;
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) betanom;
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) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = -1;
}
magma_s_vfree(&q); // frees all vectors
magma_free(d1);
magma_free(d2);
magma_free( skp );
magma_free_cpu( skp_h );
return MAGMA_SUCCESS;
} /* sbicgstab_merge */
extern "C" magma_int_t
magma_s_spmv(
float alpha,
magma_s_sparse_matrix A,
magma_s_vector x,
float beta,
magma_s_vector y,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
if ( A.memory_location != x.memory_location ||
x.memory_location != y.memory_location ) {
printf("error: linear algebra objects are not located in same memory!\n");
printf("memory locations are: %d %d %d\n",
A.memory_location, x.memory_location, y.memory_location );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_INVALID_PTR;
}
// DEV case
if ( A.memory_location == Magma_DEV ) {
if ( A.num_cols == x.num_rows && x.num_cols == 1 ) {
if ( A.storage_type == Magma_CSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
//printf("using CSR kernel for SpMV: ");
//magma_sgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
// A.dval, A.drow, A.dcol, x.dval, beta, y.dval );
//printf("done.\n");
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);
cusparseScsrmv( cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_cols, A.nnz, &alpha, descr,
A.dval, A.drow, A.dcol, x.dval, &beta, y.dval );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
//printf("using ELLPACKT kernel for SpMV: ");
magma_sgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLPACKT ) {
//printf("using ELL kernel for SpMV: ");
magma_sgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLRT ) {
//printf("using ELLRT kernel for SpMV: ");
magma_sgeellrtmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, A.drow, x.dval,
beta, y.dval, A.alignment, A.blocksize, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SELLP ) {
//printf("using SELLP kernel for SpMV: ");
magma_sgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_sgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.dval, A.num_rows, x.dval, 1, beta, y.dval,
1 );
//printf("done.\n");
}
/* else if ( A.storage_type == Magma_BCSR ) {
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
// end CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = (A.num_rows + A.blocksize-1)/A.blocksize;
int nb = (A.num_cols + A.blocksize-1)/A.blocksize;
cusparseSbsrmv( cusparseHandle, dirA,
CUSPARSE_OPERATION_NON_TRANSPOSE, mb, nb, A.numblocks,
&alpha, descr, A.dval, A.drow, A.dcol, A.blocksize, x.dval,
&beta, y.dval );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
else if ( A.num_cols < x.num_rows || x.num_cols > 1 ) {
magma_int_t num_vecs = x.num_rows / A.num_cols * x.num_cols;
if ( A.storage_type == Magma_CSR ) {
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);
if ( x.major == MagmaColMajor) {
cusparseScsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
} else if ( x.major == MagmaRowMajor) {
cusparseScsrmm2(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
CUSPARSE_OPERATION_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
}
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
if ( x.major == MagmaColMajor) {
magma_smgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_s_vector x2;
magma_svtranspose( x, &x2, queue );
magma_smgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_s_vfree(&x2, queue );
}
}
else if ( A.storage_type == Magma_ELLPACKT ) {
if ( x.major == MagmaColMajor) {
magma_smgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_s_vector x2;
magma_svtranspose( x, &x2, queue );
magma_smgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_s_vfree(&x2, queue );
}
} else if ( A.storage_type == Magma_SELLP ) {
if ( x.major == MagmaRowMajor) {
magma_smgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
}
else if ( x.major == MagmaColMajor) {
// transpose first to row major
magma_s_vector x2;
magma_svtranspose( x, &x2, queue );
magma_smgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x2.dval, beta, y.dval, queue );
magma_s_vfree(&x2, queue );
}
}/*
if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_smgemv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.dval, A.num_rows, x.dval, 1,
beta, y.dval, 1 );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else {
printf("error: CPU not yet supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- running magma_scg magma_scg_merge
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magma_s_solver_par solver_par;
solver_par.epsilon = 10e-16;
solver_par.maxiter = 1000;
solver_par.verbose = 0;
solver_par.num_eigenvalues = 0;
magma_s_preconditioner precond_par;
precond_par.solver = Magma_JACOBI;
precond_par.levels = 0;
precond_par.sweeps = 10;
int precond = 0;
int format = 0;
int version = 0;
int scale = 0;
magma_scale_t scaling = Magma_NOSCALE;
magma_s_sparse_matrix A, B, B_d;
magma_s_vector x, b;
B.blocksize = 8;
B.alignment = 8;
float one = MAGMA_S_MAKE(1.0, 0.0);
float zero = MAGMA_S_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("--precond", argv[i]) == 0 ) {
precond = atoi( argv[++i] );
switch( precond ) {
case 0: precond_par.solver = Magma_JACOBI; break;
case 1: precond_par.solver = Magma_ICC; break;
case 2: precond_par.solver = Magma_AICC; break;
}
}else if ( strcmp("--version", argv[i]) == 0 ) {
version = atoi( argv[++i] );
}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("--maxiter", argv[i]) == 0 ) {
solver_par.maxiter = atoi( argv[++i] );
} else if ( strcmp("--tol", argv[i]) == 0 ) {
sscanf( argv[++i], "%f", &solver_par.epsilon );
} else if ( strcmp("--levels", argv[i]) == 0 ) {
precond_par.levels = atoi( argv[++i] );
}else if ( strcmp("--sweeps", argv[i]) == 0 ) {
precond_par.sweeps = atoi( argv[++i] );
} else
break;
}
printf( "\n# usage: ./run_spcg"
" [ --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)"
" --maxiter %d --tol %.2e"
" --precond %d (0=Jacobi, 1=IC, 2=AIC [ --levels %d --sweeps %d]) ]"
" matrices \n\n", format, (int)B.blocksize, (int)B.alignment,
scale,
(int)solver_par.verbose,
(int)solver_par.maxiter, solver_par.epsilon,
precond, (int) precond_par.levels, (int) precond_par.sweeps );
magma_ssolverinfo_init( &solver_par, &precond_par );
while( i < argc ){
magma_s_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_smscale( &A, scaling );
magma_s_mconvert( A, &B, Magma_CSR, B.storage_type );
magma_s_mtransfer( B, &B_d, Magma_CPU, Magma_DEV );
// vectors and initial guess
magma_s_vinit( &b, Magma_DEV, A.num_cols, one );
magma_s_vinit( &x, Magma_DEV, A.num_cols, one );
magma_s_spmv( one, B_d, x, zero, b ); // b = A x
magma_s_vfree(&x);
magma_s_vinit( &x, Magma_DEV, A.num_cols, zero );
magma_s_precondsetup( B_d, b, &precond_par );
magma_spcg( B_d, b, &x, &solver_par, &precond_par );
magma_ssolverinfo( &solver_par, &precond_par );
magma_s_mfree(&B_d);
magma_s_mfree(&B);
magma_s_mfree(&A);
magma_s_vfree(&x);
magma_s_vfree(&b);
i++;
}
magma_ssolverinfo_free( &solver_par, &precond_par );
TESTING_FINALIZE();
return 0;
}
extern "C" magma_int_t
magma_scg(
magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector *x,
magma_s_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
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_s_vector r, p, q;
magma_s_vinit( &r, Magma_DEV, dofs, c_zero, queue );
magma_s_vinit( &p, Magma_DEV, dofs, c_zero, queue );
magma_s_vinit( &q, Magma_DEV, dofs, c_zero, queue );
// solver variables
float alpha, beta;
float nom, nom0, r0, betanom, betanomsq, den;
// solver setup
magma_sscal( dofs, c_zero, x->dval, 1) ; // x = 0
magma_scopy( dofs, b.dval, 1, r.dval, 1 ); // r = b
magma_scopy( dofs, b.dval, 1, p.dval, 1 ); // p = b
nom0 = betanom = magma_snrm2( dofs, r.dval, 1 );
nom = nom0 * nom0; // nom = r' * r
magma_s_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_S_REAL( magma_sdot(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_S_MAKE(nom/den, 0.);
magma_saxpy(dofs, alpha, p.dval, 1, x->dval, 1); // x = x + alpha p
magma_saxpy(dofs, -alpha, q.dval, 1, r.dval, 1); // r = r - alpha q
betanom = magma_snrm2(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_S_MAKE(betanomsq/nom, 0.); // beta = betanoms/nom
magma_sscal(dofs, beta, p.dval, 1); // p = beta*p
magma_saxpy(dofs, c_one, r.dval, 1, p.dval, 1); // p = p + r
magma_s_spmv( c_one, A, p, c_zero, q, queue ); // q = A p
den = MAGMA_S_REAL(magma_sdot(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_sresidual( 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_s_vfree(&r, queue );
magma_s_vfree(&p, queue );
magma_s_vfree(&q, queue );
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
} /* magma_scg */
magma_int_t
magma_spcg( magma_s_sparse_matrix A, magma_s_vector b, magma_s_vector *x,
magma_s_solver_par *solver_par,
magma_s_preconditioner *precond_par ){
// prepare solver feedback
solver_par->solver = Magma_PCG;
solver_par->numiter = 0;
solver_par->info = 0;
// local variables
float c_zero = MAGMA_S_ZERO, c_one = MAGMA_S_ONE;
magma_int_t dofs = A.num_rows;
// GPU workspace
magma_s_vector r, rt, p, q, h;
magma_s_vinit( &r, Magma_DEV, dofs, c_zero );
magma_s_vinit( &rt, Magma_DEV, dofs, c_zero );
magma_s_vinit( &p, Magma_DEV, dofs, c_zero );
magma_s_vinit( &q, Magma_DEV, dofs, c_zero );
magma_s_vinit( &h, Magma_DEV, dofs, c_zero );
// solver variables
float alpha, beta;
float nom, nom0, r0, gammaold, gammanew, den, res;
// solver setup
magma_sscal( dofs, c_zero, x->val, 1) ; // x = 0
magma_scopy( dofs, b.val, 1, r.val, 1 ); // r = b
// preconditioner
magma_s_applyprecond_left( A, r, &rt, precond_par );
magma_s_applyprecond_right( A, rt, &h, precond_par );
magma_scopy( dofs, h.val, 1, p.val, 1 ); // p = h
nom = MAGMA_S_REAL( magma_sdot(dofs, r.val, 1, h.val, 1) );
nom0 = magma_snrm2( dofs, r.val, 1 );
magma_s_spmv( c_one, A, p, c_zero, q ); // q = A p
den = MAGMA_S_REAL( magma_sdot(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_s_applyprecond_left( A, r, &rt, precond_par );
magma_s_applyprecond_right( A, rt, &h, precond_par );
gammanew = MAGMA_S_REAL( magma_sdot(dofs, r.val, 1, h.val, 1) );
// gn = < r,h>
if( solver_par->numiter==1 ){
magma_scopy( dofs, h.val, 1, p.val, 1 ); // p = h
}else{
beta = MAGMA_S_MAKE(gammanew/gammaold, 0.); // beta = gn/go
magma_sscal(dofs, beta, p.val, 1); // p = beta*p
magma_saxpy(dofs, c_one, h.val, 1, p.val, 1); // p = p + h
}
magma_s_spmv( c_one, A, p, c_zero, q ); // q = A p
den = MAGMA_S_REAL(magma_sdot(dofs, p.val, 1, q.val, 1));
// den = p dot q
alpha = MAGMA_S_MAKE(gammanew/den, 0.);
magma_saxpy(dofs, alpha, p.val, 1, x->val, 1); // x = x + alpha p
magma_saxpy(dofs, -alpha, q.val, 1, r.val, 1); // r = r - alpha q
gammaold = gammanew;
res = magma_snrm2( 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;
float residual;
magma_sresidual( 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_s_vfree(&r);
magma_s_vfree(&rt);
magma_s_vfree(&p);
magma_s_vfree(&q);
magma_s_vfree(&h);
return MAGMA_SUCCESS;
} /* magma_scg */
