extern "C" magma_int_t
magma_zpbicg(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_PBICG;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t dofs = A.num_rows * b.num_cols;
// workspace
magma_z_matrix r={Magma_CSR}, rt={Magma_CSR}, p={Magma_CSR}, pt={Magma_CSR},
z={Magma_CSR}, zt={Magma_CSR}, q={Magma_CSR}, y={Magma_CSR},
yt={Magma_CSR}, qt={Magma_CSR};
// need to transpose the matrix
magma_z_matrix AT={Magma_CSR}, Ah1={Magma_CSR}, Ah2={Magma_CSR};
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &rt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &pt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &qt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &y, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &yt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &zt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver variables
magmaDoubleComplex alpha, rho, beta, rho_new, ptq;
double res, nomb, nom0, r0;
// transpose the matrix
magma_zmtransfer( A, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_zmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_zmfree(&Ah1, queue );
magma_zmtransposeconjugate( Ah2, &Ah1, queue );
magma_zmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_zmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_zmfree(&Ah1, queue );
magma_zmtransfer( Ah2, &AT, Magma_CPU, Magma_DEV, queue );
magma_zmfree(&Ah2, queue );
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
res = nom0;
solver_par->init_res = nom0;
magma_zcopy( dofs, r.dval, 1, rt.dval, 1, queue ); // rr = r
rho_new = magma_zdotc( dofs, rt.dval, 1, r.dval, 1, queue ); // rho=<rr,r>
rho = alpha = MAGMA_Z_MAKE( 1.0, 0. );
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
if ( nom0 < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, r, &y, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, y, &z, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaTrans, A, rt, &yt, precond_par, queue ));
CHECK( magma_z_applyprecond_left( MagmaTrans, A, yt, &zt, precond_par, queue ));
//magma_zcopy( dofs, r.dval, 1 , y.dval, 1, queue ); // y=r
//magma_zcopy( dofs, y.dval, 1 , z.dval, 1, queue ); // z=y
//magma_zcopy( dofs, rt.dval, 1 , yt.dval, 1, queue ); // yt=rt
//magma_zcopy( dofs, yt.dval, 1 , zt.dval, 1, queue ); // yt=rt
rho= rho_new;
rho_new = magma_zdotc( dofs, rt.dval, 1, z.dval, 1, queue ); // rho=<rt,z>
if( magma_z_isnan_inf( rho_new ) ){
info = MAGMA_DIVERGENCE;
break;
}
if( solver_par->numiter==1 ){
magma_zcopy( dofs, z.dval, 1 , p.dval, 1, queue ); // yt=rt
magma_zcopy( dofs, zt.dval, 1 , pt.dval, 1, queue ); // zt=yt
} else {
beta = rho_new/rho;
magma_zscal( dofs, beta, p.dval, 1, queue ); // p = beta*p
magma_zaxpy( dofs, c_one , z.dval, 1 , p.dval, 1, queue ); // p = z+beta*p
magma_zscal( dofs, MAGMA_Z_CONJ(beta), pt.dval, 1, queue ); // pt = beta*pt
magma_zaxpy( dofs, c_one , zt.dval, 1 , pt.dval, 1, queue ); // pt = zt+beta*pt
}
CHECK( magma_z_spmv( c_one, A, p, c_zero, q, queue )); // v = Ap
CHECK( magma_z_spmv( c_one, AT, pt, c_zero, qt, queue )); // v = Ap
solver_par->spmv_count++;
solver_par->spmv_count++;
ptq = magma_zdotc( dofs, pt.dval, 1, q.dval, 1, queue );
alpha = rho_new /ptq;
magma_zaxpy( dofs, alpha, p.dval, 1 , x->dval, 1, queue ); // x=x+alpha*p
magma_zaxpy( dofs, c_neg_one * alpha, q.dval, 1 , r.dval, 1, queue ); // r=r+alpha*q
magma_zaxpy( dofs, c_neg_one * MAGMA_Z_CONJ(alpha), qt.dval, 1 , rt.dval, 1, queue ); // r=r+alpha*q
res = magma_dznrm2( dofs, r.dval, 1, queue );
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&rt, queue );
magma_zmfree(&p, queue );
magma_zmfree(&pt, queue );
magma_zmfree(&q, queue );
magma_zmfree(&qt, queue );
magma_zmfree(&y, queue );
magma_zmfree(&yt, queue );
magma_zmfree(&z, queue );
magma_zmfree(&zt, queue );
magma_zmfree(&AT, queue );
magma_zmfree(&Ah1, queue );
magma_zmfree(&Ah2, queue );
solver_par->info = info;
return info;
} /* magma_zpbicg */
extern "C" magma_int_t
magma_zqmr_merge(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_QMRMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// local variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
// solver variables
double nom0, r0, res=0, nomb;
magmaDoubleComplex rho = c_one, rho1 = c_one, eta = -c_one , pds = c_one,
thet = c_one, thet1 = c_one, epsilon = c_one,
beta = c_one, delta = c_one, pde = c_one, rde = c_one,
gamm = c_one, gamm1 = c_one, psi = c_one;
magma_int_t dofs = A.num_rows* b.num_cols;
// need to transpose the matrix
magma_z_matrix AT={Magma_CSR}, Ah1={Magma_CSR}, Ah2={Magma_CSR};
// GPU workspace
magma_z_matrix r={Magma_CSR}, r_tld={Magma_CSR},
v={Magma_CSR}, w={Magma_CSR}, wt={Magma_CSR},
d={Magma_CSR}, s={Magma_CSR}, z={Magma_CSR}, q={Magma_CSR},
p={Magma_CSR}, pt={Magma_CSR}, y={Magma_CSR};
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &r_tld, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &w, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &wt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &d, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &s, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &pt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &y, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
solver_par->init_res = nom0;
magma_zcopy( dofs, r.dval, 1, r_tld.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, y.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, v.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, wt.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, z.dval, 1, queue );
// transpose the matrix
magma_zmtransfer( A, &Ah1, Magma_DEV, Magma_CPU, queue );
magma_zmconvert( Ah1, &Ah2, A.storage_type, Magma_CSR, queue );
magma_zmfree(&Ah1, queue );
magma_zmtransposeconjugate( Ah2, &Ah1, queue );
magma_zmfree(&Ah2, queue );
Ah2.blocksize = A.blocksize;
Ah2.alignment = A.alignment;
magma_zmconvert( Ah1, &Ah2, Magma_CSR, A.storage_type, queue );
magma_zmfree(&Ah1, queue );
magma_zmtransfer( Ah2, &AT, Magma_CPU, Magma_DEV, queue );
magma_zmfree(&Ah2, queue );
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nom0;
solver_par->timing[0] = 0.0;
}
if ( nom0 < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
psi = magma_zsqrt( magma_zdotc( dofs, z.dval, 1, z.dval, 1, queue ));
rho = magma_zsqrt( magma_zdotc( dofs, y.dval, 1, y.dval, 1, queue ));
// v = y / rho
// y = y / rho
// w = wt / psi
// z = z / psi
magma_zqmr_1(
r.num_rows,
r.num_cols,
rho,
psi,
y.dval,
z.dval,
v.dval,
w.dval,
queue );
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
if( magma_z_isnan_inf( rho ) || magma_z_isnan_inf( psi ) ){
info = MAGMA_DIVERGENCE;
break;
}
// delta = z' * y;
delta = magma_zdotc( dofs, z.dval, 1, y.dval, 1, queue );
if( magma_z_isnan_inf( delta ) ){
info = MAGMA_DIVERGENCE;
break;
}
// no precond: yt = y, zt = z
//magma_zcopy( dofs, y.dval, 1, yt.dval, 1 );
//magma_zcopy( dofs, z.dval, 1, zt.dval, 1 );
if( solver_par->numiter == 1 ){
// p = y;
// q = z;
magma_zcopy( dofs, y.dval, 1, p.dval, 1, queue );
magma_zcopy( dofs, z.dval, 1, q.dval, 1, queue );
}
else{
pde = psi * delta / epsilon;
rde = rho * MAGMA_Z_CONJ(delta/epsilon);
// p = y - pde * p
// q = z - rde * q
magma_zqmr_2(
r.num_rows,
r.num_cols,
pde,
rde,
y.dval,
z.dval,
p.dval,
q.dval,
queue );
}
if( magma_z_isnan_inf( rho ) || magma_z_isnan_inf( psi ) ){
info = MAGMA_DIVERGENCE;
break;
}
CHECK( magma_z_spmv( c_one, A, p, c_zero, pt, queue ));
solver_par->spmv_count++;
// epsilon = q' * pt;
epsilon = magma_zdotc( dofs, q.dval, 1, pt.dval, 1, queue );
beta = epsilon / delta;
if( magma_z_isnan_inf( epsilon ) || magma_z_isnan_inf( beta ) ){
info = MAGMA_DIVERGENCE;
break;
}
// v = pt - beta * v
// y = v
magma_zqmr_3(
r.num_rows,
r.num_cols,
beta,
pt.dval,
v.dval,
y.dval,
queue );
rho1 = rho;
// rho = norm(y);
rho = magma_zsqrt( magma_zdotc( dofs, y.dval, 1, y.dval, 1, queue ));
// wt = A' * q - beta' * w;
CHECK( magma_z_spmv( c_one, AT, q, c_zero, wt, queue ));
solver_par->spmv_count++;
magma_zaxpy( dofs, - MAGMA_Z_CONJ( beta ), w.dval, 1, wt.dval, 1, queue );
// no precond: z = wt
magma_zcopy( dofs, wt.dval, 1, z.dval, 1, queue );
thet1 = thet;
thet = rho / (gamm * MAGMA_Z_MAKE( MAGMA_Z_ABS(beta), 0.0 ));
gamm1 = gamm;
gamm = c_one / magma_zsqrt(c_one + thet*thet);
eta = - eta * rho1 * gamm * gamm / (beta * gamm1 * gamm1);
if( magma_z_isnan_inf( thet ) || magma_z_isnan_inf( gamm ) || magma_z_isnan_inf( eta ) ){
info = MAGMA_DIVERGENCE;
break;
}
if( solver_par->numiter == 1 ){
// d = eta * p + pds * d;
// s = eta * pt + pds * d;
// x = x + d;
// r = r - s;
magma_zqmr_4(
r.num_rows,
r.num_cols,
eta,
p.dval,
pt.dval,
d.dval,
s.dval,
x->dval,
r.dval,
queue );
}
else{
pds = (thet1 * gamm) * (thet1 * gamm);
// d = eta * p + pds * d;
// s = eta * pt + pds * d;
// x = x + d;
// r = r - s;
magma_zqmr_5(
r.num_rows,
r.num_cols,
eta,
pds,
p.dval,
pt.dval,
d.dval,
s.dval,
x->dval,
r.dval,
queue );
}
// psi = norm(z);
psi = magma_zsqrt( magma_zdotc( dofs, z.dval, 1, z.dval, 1, queue ) );
res = magma_dznrm2( dofs, r.dval, 1, queue );
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
// v = y / rho
// y = y / rho
// w = wt / psi
// z = z / psi
magma_zqmr_1(
r.num_rows,
r.num_cols,
rho,
psi,
y.dval,
z.dval,
v.dval,
w.dval,
queue );
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter && info == MAGMA_SUCCESS ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == c_zero ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&r_tld, queue );
magma_zmfree(&v, queue );
magma_zmfree(&w, queue );
magma_zmfree(&wt, queue );
magma_zmfree(&d, queue );
magma_zmfree(&s, queue );
magma_zmfree(&z, queue );
magma_zmfree(&q, queue );
magma_zmfree(&p, queue );
magma_zmfree(&pt, queue );
magma_zmfree(&y, queue );
magma_zmfree(&AT, queue );
magma_zmfree(&Ah1, queue );
magma_zmfree(&Ah2, queue );
solver_par->info = info;
return info;
} /* magma_zqmr_merge */
magma_int_t
magma_zorderstatistics(
magmaDoubleComplex *val,
magma_int_t length,
magma_int_t k,
magma_int_t r,
magmaDoubleComplex *element,
magma_queue_t queue )
{
magma_int_t info = 0;
magma_int_t i, st;
magmaDoubleComplex tmp;
if( r == 0 ){
for ( st = i = 0; i < length - 1; i++ ) {
if ( magma_z_isnan_inf( val[i]) ) {
printf("error: array contains %f + %fi.\n", MAGMA_Z_REAL(val[i]), MAGMA_Z_IMAG(val[i]) );
info = MAGMA_ERR_NAN;
goto cleanup;
}
if ( MAGMA_Z_ABS(val[i]) > MAGMA_Z_ABS(val[length-1]) ){
continue;
}
SWAP(i, st);
st++;
}
SWAP(length-1, st);
if ( k == st ){
*element = val[st];
}
else if ( st > k ) {
CHECK( magma_zorderstatistics( val, st, k, r, element, queue ));
}
else {
CHECK( magma_zorderstatistics( val+st, length-st, k-st, r, element, queue ));
}
} else {
for ( st = i = 0; i < length - 1; i++ ) {
if ( magma_z_isnan_inf( val[i]) ) {
printf("error: array contains %f + %fi.\n", MAGMA_Z_REAL(val[i]), MAGMA_Z_IMAG(val[i]) );
info = MAGMA_ERR_NAN;
goto cleanup;
}
if ( MAGMA_Z_ABS(val[i]) < MAGMA_Z_ABS(val[length-1]) ){
continue;
}
SWAP(i, st);
st++;
}
SWAP(length-1, st);
if ( k == st ){
*element = val[st];
}
else if ( st > k ) {
CHECK( magma_zorderstatistics( val, st, k, r, element, queue ));
}
else {
CHECK( magma_zorderstatistics( val+st, length-st, k-st, r, element, queue ));
}
}
cleanup:
return info;
}
extern "C" magma_int_t
magma_zpbicgstab_merge(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_BICGSTAB;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows * b.num_cols;
// workspace
magma_z_matrix r={Magma_CSR}, rr={Magma_CSR}, p={Magma_CSR}, v={Magma_CSR},
z={Magma_CSR}, y={Magma_CSR}, ms={Magma_CSR}, mt={Magma_CSR},
s={Magma_CSR}, t={Magma_CSR}, d1={Magma_CSR}, d2={Magma_CSR};
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &rr,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &s, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &t, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &ms,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &mt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &y, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &d1, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &d2, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver variables
magmaDoubleComplex alpha, beta, omega, rho_old, rho_new;
double nom, betanom, nom0, r0, res, nomb;
res=0;
//double den;
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
magma_zcopy( dofs, r.dval, 1, rr.dval, 1, queue ); // rr = r
betanom = nom0;
nom = nom0*nom0;
rho_new = magma_zdotc( dofs, r.dval, 1, r.dval, 1, queue ); // rho=<rr,r>
rho_old = omega = alpha = MAGMA_Z_MAKE( 1.0, 0. );
solver_par->init_res = nom0;
CHECK( magma_z_spmv( c_one, A, r, c_zero, v, queue )); // z = A r
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = nom0;
solver_par->timing[0] = 0.0;
}
if ( nom < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
//Chronometry
real_Double_t tempo1, tempo2, tempop1, tempop2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
rho_old = rho_new; // rho_old=rho
rho_new = magma_zdotc( dofs, rr.dval, 1, r.dval, 1, queue ); // rho=<rr,r>
beta = rho_new/rho_old * alpha/omega; // beta=rho/rho_old *alpha/omega
if( magma_z_isnan_inf( beta ) ){
info = MAGMA_DIVERGENCE;
break;
}
// p = r + beta * ( p - omega * v )
magma_zbicgstab_1(
r.num_rows,
r.num_cols,
beta,
omega,
r.dval,
v.dval,
p.dval,
queue );
// preconditioner
tempop1 = magma_sync_wtime( queue );
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, p, &mt, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, mt, &y, precond_par, queue ));
tempop2 = magma_sync_wtime( queue );
precond_par->runtime += tempop2-tempop1;
CHECK( magma_z_spmv( c_one, A, y, c_zero, v, queue )); // v = Ap
solver_par->spmv_count++;
//alpha = rho_new / tmpval;
alpha = rho_new /magma_zdotc( dofs, rr.dval, 1, v.dval, 1, queue );
if( magma_z_isnan_inf( alpha ) ){
info = MAGMA_DIVERGENCE;
break;
}
// s = r - alpha v
magma_zbicgstab_2(
r.num_rows,
r.num_cols,
alpha,
r.dval,
v.dval,
s.dval,
queue );
// preconditioner
tempop1 = magma_sync_wtime( queue );
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, s, &ms, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, ms, &z, precond_par, queue ));
tempop2 = magma_sync_wtime( queue );
precond_par->runtime += tempop2-tempop1;
CHECK( magma_z_spmv( c_one, A, z, c_zero, t, queue )); // t=As
solver_par->spmv_count++;
omega = magma_zdotc( dofs, t.dval, 1, s.dval, 1, queue ) // omega = <s,t>/<t,t>
/ magma_zdotc( dofs, t.dval, 1, t.dval, 1, queue );
// x = x + alpha * y + omega * z
// r = s - omega * t
magma_zbicgstab_4(
r.num_rows,
r.num_cols,
alpha,
omega,
y.dval,
z.dval,
s.dval,
t.dval,
x->dval,
r.dval,
queue );
res = betanom = magma_dznrm2( dofs, r.dval, 1, queue );
nom = betanom*betanom;
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r, &residual, queue));
solver_par->final_res = residual;
solver_par->iter_res = res;
if ( solver_par->numiter < solver_par->maxiter && info == MAGMA_SUCCESS ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==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;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&rr, queue );
magma_zmfree(&p, queue );
magma_zmfree(&v, queue );
magma_zmfree(&s, queue );
magma_zmfree(&y, queue );
magma_zmfree(&z, queue );
magma_zmfree(&t, queue );
magma_zmfree(&ms, queue );
magma_zmfree(&mt, queue );
magma_zmfree(&d1, queue );
magma_zmfree(&d2, queue );
solver_par->info = info;
return info;
} /* magma_zbicgstab_merge */
extern "C" magma_int_t
magma_zptfqmr_merge(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_TFQMRMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// local variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
// solver variables
double nom0, r0, res=0, nomb;
magmaDoubleComplex rho = c_one, rho_l = c_one, eta = c_zero , c = c_zero ,
theta = c_zero , tau = c_zero, alpha = c_one, beta = c_zero,
sigma = c_zero;
magma_int_t dofs = A.num_rows* b.num_cols;
// GPU workspace
magma_z_matrix r={Magma_CSR}, r_tld={Magma_CSR}, pu_m={Magma_CSR},
d={Magma_CSR}, w={Magma_CSR}, v={Magma_CSR}, t={Magma_CSR},
u_mp1={Magma_CSR}, u_m={Magma_CSR}, Au={Magma_CSR},
Ad={Magma_CSR}, Au_new={Magma_CSR};
CHECK( magma_zvinit( &t, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &u_mp1,Magma_DEV, A.num_rows, b.num_cols, c_one, queue ));
CHECK( magma_zvinit( &r_tld,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &u_m, Magma_DEV, A.num_rows, b.num_cols, c_one, queue ));
CHECK( magma_zvinit( &pu_m, Magma_DEV, A.num_rows, b.num_cols, c_one, queue ));
CHECK( magma_zvinit( &v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &d, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &w, Magma_DEV, A.num_rows, b.num_cols, c_one, queue ));
CHECK( magma_zvinit( &Ad, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Au_new, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Au, Magma_DEV, A.num_rows, b.num_cols, c_one, queue ));
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
solver_par->init_res = nom0;
magma_zcopy( dofs, r.dval, 1, r_tld.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, w.dval, 1, queue );
magma_zcopy( dofs, r.dval, 1, u_m.dval, 1, queue );
// preconditioner
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, u_m, &t, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, t, &pu_m, precond_par, queue ));
CHECK( magma_z_spmv( c_one, A, pu_m, c_zero, v, queue )); // v = A u
magma_zcopy( dofs, v.dval, 1, Au.dval, 1, queue );
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nom0;
solver_par->timing[0] = 0.0;
}
if ( nom0 < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
tau = magma_zsqrt( magma_zdotc( dofs, r.dval, 1, r_tld.dval, 1, queue) );
rho = magma_zdotc( dofs, r.dval, 1, r_tld.dval, 1, queue );
rho_l = rho;
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
// do this every iteration as unrolled
alpha = rho / magma_zdotc( dofs, v.dval, 1, r_tld.dval, 1, queue );
sigma = theta * theta / alpha * eta;
magma_ztfqmr_1(
r.num_rows,
r.num_cols,
alpha,
sigma,
v.dval,
Au.dval,
u_m.dval,
pu_m.dval,
u_mp1.dval,
w.dval,
d.dval,
Ad.dval,
queue );
theta = magma_zsqrt( magma_zdotc(dofs, w.dval, 1, w.dval, 1, queue) ) / tau;
c = c_one / magma_zsqrt( c_one + theta*theta );
tau = tau * theta *c;
eta = c * c * alpha;
sigma = theta * theta / alpha * eta;
if ( magma_z_isnan_inf( theta ) ||
magma_z_isnan_inf( c ) ||
magma_z_isnan_inf( tau ) ||
magma_z_isnan_inf( eta ) ||
magma_z_isnan_inf( sigma ) )
{
info = MAGMA_DIVERGENCE;
break;
}
magma_ztfqmr_2(
r.num_rows,
r.num_cols,
eta,
d.dval,
Ad.dval,
x->dval,
r.dval,
queue );
res = magma_dznrm2( dofs, r.dval, 1, queue );
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
info = MAGMA_SUCCESS;
break;
}
// preconditioner
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, u_mp1, &t, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, t, &pu_m, precond_par, queue ));
CHECK( magma_z_spmv( c_one, A, pu_m, c_zero, Au_new, queue )); // Au_new = A u_mp1
solver_par->spmv_count++;
magma_zcopy( dofs, Au_new.dval, 1, Au.dval, 1, queue );
magma_zcopy( dofs, u_mp1.dval, 1, u_m.dval, 1, queue );
// here starts the second part of the loop #################################
magma_ztfqmr_5(
r.num_rows,
r.num_cols,
alpha,
sigma,
v.dval,
Au.dval,
pu_m.dval,
w.dval,
d.dval,
Ad.dval,
queue );
sigma = theta * theta / alpha * eta;
theta = magma_zsqrt( magma_zdotc(dofs, w.dval, 1, w.dval, 1, queue) ) / tau;
c = c_one / magma_zsqrt( c_one + theta*theta );
tau = tau * theta *c;
eta = c * c * alpha;
magma_ztfqmr_2(
r.num_rows,
r.num_cols,
eta,
d.dval,
Ad.dval,
x->dval,
r.dval,
queue );
res = magma_dznrm2( dofs, r.dval, 1, queue );
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
info = MAGMA_SUCCESS;
break;
}
rho = magma_zdotc( dofs, w.dval, 1, r_tld.dval, 1, queue );
beta = rho / rho_l;
rho_l = rho;
magma_ztfqmr_3(
r.num_rows,
r.num_cols,
beta,
w.dval,
u_m.dval,
u_mp1.dval,
queue );
// preconditioner
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, u_mp1, &t, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, t, &pu_m, precond_par, queue ));
CHECK( magma_z_spmv( c_one, A, pu_m, c_zero, Au_new, queue )); // Au_new = A pu_m
solver_par->spmv_count++;
magma_ztfqmr_4(
r.num_rows,
r.num_cols,
beta,
Au_new.dval,
v.dval,
Au.dval,
queue );
magma_zcopy( dofs, u_mp1.dval, 1, u_m.dval, 1, queue );
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->rtol*solver_par->init_res ||
solver_par->iter_res < solver_par->atol ) {
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&r_tld, queue );
magma_zmfree(&d, queue );
magma_zmfree(&w, queue );
magma_zmfree(&v, queue );
magma_zmfree(&u_m, queue );
magma_zmfree(&u_mp1, queue );
magma_zmfree(&pu_m, queue );
magma_zmfree(&d, queue );
magma_zmfree(&t, queue );
magma_zmfree(&Au, queue );
magma_zmfree(&Au_new, queue );
magma_zmfree(&Ad, queue );
solver_par->info = info;
return info;
} /* magma_zptfqmr_merge */
