extern "C" magma_int_t
magma_zresidual(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix x,
double *res,
magma_queue_t queue )
{
magma_int_t info = 0;
// set queue for old dense routines
magma_queue_t orig_queue=NULL;
magmablasGetKernelStream( &orig_queue );
// some useful variables
magmaDoubleComplex zero = MAGMA_Z_ZERO, one = MAGMA_Z_ONE,
mone = MAGMA_Z_NEG_ONE;
magma_int_t dofs = A.num_rows;
magma_int_t num_vecs = b.num_rows*b.num_cols/A.num_rows;
magma_z_matrix r={Magma_CSR};
if ( A.num_rows == b.num_rows ) {
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, zero, queue ));
CHECK( magma_z_spmv( one, A, x, zero, r, queue )); // r = A x
magma_zaxpy(dofs, mone, b.dval, 1, r.dval, 1); // r = r - b
*res = magma_dznrm2(dofs, r.dval, 1); // res = ||r||
// /magma_dznrm2(dofs, b.dval, 1); /||b||
//printf( "relative residual: %e\n", *res );
} else if ((b.num_rows*b.num_cols)%A.num_rows== 0 ) {
CHECK( magma_zvinit( &r, Magma_DEV, b.num_rows,b.num_cols, zero, queue ));
CHECK( magma_z_spmv( one, A, x, zero, r, queue )); // r = A x
for( magma_int_t i=0; i<num_vecs; i++) {
magma_zaxpy(dofs, mone, b(i), 1, r(i), 1); // r = r - b
res[i] = magma_dznrm2(dofs, r(i), 1); // res = ||r||
}
// /magma_dznrm2(dofs, b.dval, 1); /||b||
//printf( "relative residual: %e\n", *res );
} else {
printf("error: dimensions do not match.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
magma_zmfree(&r, queue );
magmablasSetKernelStream( orig_queue );
return info;
}
extern "C" magma_int_t
magma_zresidual_slice(
magma_int_t start, magma_int_t end,
magma_z_matrix A, magma_z_matrix b, magma_z_matrix x,
double *res,
magma_queue_t queue )
{
magma_int_t info = 0;
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
// some useful variables
magma_int_t dofs = A.num_rows;
magma_int_t num_vecs = b.num_rows*b.num_cols/A.num_rows;
magma_z_matrix r = {Magma_CSR};
if ( A.num_rows == b.num_rows ) {
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_z_spmv( c_one, A, x, c_zero, r, queue )); // r = A x
magma_zaxpy( dofs, c_neg_one, b.dval, 1, r.dval, 1, queue ); // r = r - b
*res = magma_dznrm2( end-start, r.dval+start, 1, queue ); // res = ||r(start:end)||
} else if ((b.num_rows*b.num_cols)%A.num_rows == 0 ) {
CHECK( magma_zvinit( &r, Magma_DEV, b.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_z_spmv( c_one, A, x, c_zero, r, queue )); // r = A x
for( magma_int_t i=0; i < num_vecs; i++) {
magma_zaxpy( dofs, c_neg_one, b(i), 1, r(i), 1, queue ); // r = r - b
res[i] = magma_dznrm2( end-start, r(i)+start, 1, queue ); // res = ||r(start:end)||
}
} else {
printf("error: dimensions do not match.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
}
cleanup:
magma_zmfree( &r, queue );
return info;
}
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 */
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 */
extern "C" magma_int_t
magma_zpidr_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_PIDRMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
solver_par->init_res = 0.0;
solver_par->final_res = 0.0;
solver_par->iter_res = 0.0;
solver_par->runtime = 0.0;
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_n_one = MAGMA_Z_NEG_ONE;
// internal user parameters
const magma_int_t smoothing = 1; // 0 = disable, 1 = enable
const double angle = 0.7; // [0-1]
// local variables
magma_int_t iseed[4] = {0, 0, 0, 1};
magma_int_t dof;
magma_int_t s;
magma_int_t distr;
magma_int_t k, i, sk;
magma_int_t innerflag;
magma_int_t ldd;
double residual;
double nrm;
double nrmb;
double nrmr;
double nrmt;
double rho;
magmaDoubleComplex om;
magmaDoubleComplex gamma;
magmaDoubleComplex fk;
// matrices and vectors
magma_z_matrix dxs = {Magma_CSR};
magma_z_matrix dr = {Magma_CSR}, drs = {Magma_CSR};
magma_z_matrix dP = {Magma_CSR}, dP1 = {Magma_CSR};
magma_z_matrix dG = {Magma_CSR}, dGcol = {Magma_CSR};
magma_z_matrix dU = {Magma_CSR};
magma_z_matrix dM = {Magma_CSR}, hMdiag = {Magma_CSR};
magma_z_matrix df = {Magma_CSR};
magma_z_matrix dt = {Magma_CSR}, dtt = {Magma_CSR};
magma_z_matrix dc = {Magma_CSR};
magma_z_matrix dv = {Magma_CSR};
magma_z_matrix dlu = {Magma_CSR};
magma_z_matrix dskp = {Magma_CSR}, hskp = {Magma_CSR};
magma_z_matrix dalpha = {Magma_CSR}, halpha = {Magma_CSR};
magma_z_matrix dbeta = {Magma_CSR}, hbeta = {Magma_CSR};
magmaDoubleComplex *d1 = NULL, *d2 = NULL;
// chronometry
real_Double_t tempo1, tempo2;
// initial s space
// TODO: add option for 's' (shadow space number)
// Hack: uses '--restart' option as the shadow space number.
// This is not a good idea because the default value of restart option is used to detect
// if the user provided a custom restart. This means that if the default restart value
// is changed then the code will think it was the user (unless the default value is
// also updated in the 'if' statement below.
s = 1;
if ( solver_par->restart != 50 ) {
if ( solver_par->restart > A.num_cols ) {
s = A.num_cols;
} else {
s = solver_par->restart;
}
}
solver_par->restart = s;
// set max iterations
solver_par->maxiter = min( 2 * A.num_cols, solver_par->maxiter );
// check if matrix A is square
if ( A.num_rows != A.num_cols ) {
//printf("Matrix A is not square.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
// |b|
nrmb = magma_dznrm2( b.num_rows, b.dval, 1, queue );
if ( nrmb == 0.0 ) {
magma_zscal( x->num_rows, MAGMA_Z_ZERO, x->dval, 1, queue );
info = MAGMA_SUCCESS;
goto cleanup;
}
// t = 0
// make t twice as large to contain both, dt and dr
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_zvinit( &dt, Magma_DEV, ldd, 2, c_zero, queue ));
dt.num_rows = b.num_rows;
dt.num_cols = 1;
dt.nnz = dt.num_rows;
// redirect the dr.dval to the second part of dt
CHECK( magma_zvinit( &dr, Magma_DEV, b.num_rows, 1, c_zero, queue ));
magma_free( dr.dval );
dr.dval = dt.dval + ldd;
// r = b - A x
CHECK( magma_zresidualvec( A, b, *x, &dr, &nrmr, queue ));
// |r|
solver_par->init_res = nrmr;
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nrmr;
}
// check if initial is guess good enough
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// P = randn(n, s)
// P = ortho(P)
//---------------------------------------
// P = 0.0
CHECK( magma_zvinit( &dP, Magma_CPU, A.num_cols, s, c_zero, queue ));
// P = randn(n, s)
distr = 3; // 1 = unif (0,1), 2 = unif (-1,1), 3 = normal (0,1)
dof = dP.num_rows * dP.num_cols;
lapackf77_zlarnv( &distr, iseed, &dof, dP.val );
// transfer P to device
CHECK( magma_zmtransfer( dP, &dP1, Magma_CPU, Magma_DEV, queue ));
magma_zmfree( &dP, queue );
// P = ortho(P1)
if ( dP1.num_cols > 1 ) {
// P = magma_zqr(P1), QR factorization
CHECK( magma_zqr( dP1.num_rows, dP1.num_cols, dP1, dP1.ld, &dP, NULL, queue ));
} else {
// P = P1 / |P1|
nrm = magma_dznrm2( dof, dP1.dval, 1, queue );
nrm = 1.0 / nrm;
magma_zdscal( dof, nrm, dP1.dval, 1, queue );
CHECK( magma_zmtransfer( dP1, &dP, Magma_DEV, Magma_DEV, queue ));
}
magma_zmfree( &dP1, queue );
//---------------------------------------
// allocate memory for the scalar products
CHECK( magma_zvinit( &hskp, Magma_CPU, 4, 1, c_zero, queue ));
CHECK( magma_zvinit( &dskp, Magma_DEV, 4, 1, c_zero, queue ));
CHECK( magma_zvinit( &halpha, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &dalpha, Magma_DEV, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &hbeta, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &dbeta, Magma_DEV, s, 1, c_zero, queue ));
// workspace for merged dot product
CHECK( magma_zmalloc( &d1, max(2, s) * b.num_rows ));
CHECK( magma_zmalloc( &d2, max(2, s) * b.num_rows ));
// smoothing enabled
if ( smoothing > 0 ) {
// set smoothing solution vector
CHECK( magma_zmtransfer( *x, &dxs, Magma_DEV, Magma_DEV, queue ));
// tt = 0
// make tt twice as large to contain both, dtt and drs
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_zvinit( &dtt, Magma_DEV, ldd, 2, c_zero, queue ));
dtt.num_rows = dr.num_rows;
dtt.num_cols = 1;
dtt.nnz = dtt.num_rows;
// redirect the drs.dval to the second part of dtt
CHECK( magma_zvinit( &drs, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
magma_free( drs.dval );
drs.dval = dtt.dval + ldd;
// set smoothing residual vector
magma_zcopyvector( dr.num_rows, dr.dval, 1, drs.dval, 1, queue );
}
// G(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_rows, 32 );
CHECK( magma_zvinit( &dG, Magma_DEV, ldd, s, c_zero, queue ));
dG.num_rows = A.num_rows;
} else {
CHECK( magma_zvinit( &dG, Magma_DEV, A.num_rows, s, c_zero, queue ));
}
// dGcol represents a single column of dG, array pointer is set inside loop
CHECK( magma_zvinit( &dGcol, Magma_DEV, dG.num_rows, 1, c_zero, queue ));
magma_free( dGcol.dval );
// U(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_cols, 32 );
CHECK( magma_zvinit( &dU, Magma_DEV, ldd, s, c_zero, queue ));
dU.num_rows = A.num_cols;
} else {
CHECK( magma_zvinit( &dU, Magma_DEV, A.num_cols, s, c_zero, queue ));
}
// M(s,s) = I
CHECK( magma_zvinit( &dM, Magma_DEV, s, s, c_zero, queue ));
CHECK( magma_zvinit( &hMdiag, Magma_CPU, s, 1, c_zero, queue ));
magmablas_zlaset( MagmaFull, dM.num_rows, dM.num_cols, c_zero, c_one, dM.dval, dM.ld, queue );
// f = 0
CHECK( magma_zvinit( &df, Magma_DEV, dP.num_cols, 1, c_zero, queue ));
// c = 0
CHECK( magma_zvinit( &dc, Magma_DEV, dM.num_cols, 1, c_zero, queue ));
// v = 0
CHECK( magma_zvinit( &dv, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
// lu = 0
CHECK( magma_zvinit( &dlu, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
//--------------START TIME---------------
// chronometry
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->timing[0] = 0.0;
}
om = MAGMA_Z_ONE;
innerflag = 0;
// start iteration
do
{
solver_par->numiter++;
// new RHS for small systems
// f = P' r
magma_zgemvmdot_shfl( dP.num_rows, dP.num_cols, dP.dval, dr.dval, d1, d2, df.dval, queue );
// shadow space loop
for ( k = 0; k < s; ++k ) {
sk = s - k;
// c(k:s) = M(k:s,k:s) \ f(k:s)
magma_zcopyvector( sk, &df.dval[k], 1, &dc.dval[k], 1, queue );
magma_ztrsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, sk, &dM.dval[k*dM.ld+k], dM.ld, &dc.dval[k], 1, queue );
// v = r - G(:,k:s) c(k:s)
magma_zcopyvector( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dG.num_rows, sk, c_n_one, &dG.dval[k*dG.ld], dG.ld, &dc.dval[k], 1, c_one, dv.dval, 1, queue );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// U(:,k) = om * v + U(:,k:s) c(k:s)
magmablas_zgemv( MagmaNoTrans, dU.num_rows, sk, c_one, &dU.dval[k*dU.ld], dU.ld, &dc.dval[k], 1, om, dv.dval, 1, queue );
magma_zcopyvector( dU.num_rows, dv.dval, 1, &dU.dval[k*dU.ld], 1, queue );
// G(:,k) = A U(:,k)
dGcol.dval = dG.dval + k * dG.ld;
CHECK( magma_z_spmv( c_one, A, dv, c_zero, dGcol, queue ));
solver_par->spmv_count++;
// bi-orthogonalize the new basis vectors
for ( i = 0; i < k; ++i ) {
// alpha = P(:,i)' G(:,k)
halpha.val[i] = magma_zdotc( dP.num_rows, &dP.dval[i*dP.ld], 1, &dG.dval[k*dG.ld], 1, queue );
// alpha = alpha / M(i,i)
halpha.val[i] = halpha.val[i] / hMdiag.val[i];
// G(:,k) = G(:,k) - alpha * G(:,i)
magma_zaxpy( dG.num_rows, -halpha.val[i], &dG.dval[i*dG.ld], 1, &dG.dval[k*dG.ld], 1, queue );
}
// non-first s iteration
if ( k > 0 ) {
// U update outside of loop using GEMV
// U(:,k) = U(:,k) - U(:,1:k) * alpha(1:k)
magma_zsetvector( k, halpha.val, 1, dalpha.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dU.num_rows, k, c_n_one, dU.dval, dU.ld, dalpha.dval, 1, c_one, &dU.dval[k*dU.ld], 1, queue );
}
// new column of M = P'G, first k-1 entries are zero
// M(k:s,k) = P(:,k:s)' G(:,k)
magma_zgemvmdot_shfl( dP.num_rows, sk, &dP.dval[k*dP.ld], &dG.dval[k*dG.ld], d1, d2, &dM.dval[k*dM.ld+k], queue );
magma_zgetvector( 1, &dM.dval[k*dM.ld+k], 1, &hMdiag.val[k], 1, queue );
// check M(k,k) == 0
if ( MAGMA_Z_EQUAL(hMdiag.val[k], MAGMA_Z_ZERO) ) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// beta = f(k) / M(k,k)
magma_zgetvector( 1, &df.dval[k], 1, &fk, 1, queue );
hbeta.val[k] = fk / hMdiag.val[k];
// check for nan
if ( magma_z_isnan( hbeta.val[k] ) || magma_z_isinf( hbeta.val[k] )) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// r = r - beta * G(:,k)
magma_zaxpy( dr.num_rows, -hbeta.val[k], &dG.dval[k*dG.ld], 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
nrmr = magma_dznrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// x = x + beta * U(:,k)
magma_zaxpy( x->num_rows, hbeta.val[k], &dU.dval[k*dU.ld], 1, x->dval, 1, queue );
// smoothing operation
//---------------------------------------
// t = rs - r
magma_zidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queue );
// t't
// t'rs
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queue ));
magma_zgetvector( 2, &dskp.dval[2], 1, &hskp.val[2], 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = hskp.val[3] / hskp.val[2];
// rs = rs - gamma * (rs - r)
magma_zaxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_zidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queue );
// |rs|
nrmr = magma_dznrm2( drs.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence or iteration limit
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
s = k + 1; // for the x-update outside the loop
innerflag = 2;
info = MAGMA_SUCCESS;
break;
}
// non-last s iteration
if ( (k + 1) < s ) {
// f(k+1:s) = f(k+1:s) - beta * M(k+1:s,k)
magma_zaxpy( sk-1, -hbeta.val[k], &dM.dval[k*dM.ld+(k+1)], 1, &df.dval[k+1], 1, queue );
}
}
// smoothing disabled
if ( smoothing <= 0 && innerflag != 1 ) {
// update solution approximation x
// x = x + U(:,1:s) * beta(1:s)
magma_zsetvector( s, hbeta.val, 1, dbeta.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dU.num_rows, s, c_one, dU.dval, dU.ld, dbeta.dval, 1, c_one, x->dval, 1, queue );
}
// check convergence or iteration limit or invalid result of inner loop
if ( innerflag > 0 ) {
break;
}
// v = r
magma_zcopy( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// t = A v
CHECK( magma_z_spmv( c_one, A, dv, c_zero, dt, queue ));
solver_par->spmv_count++;
// computation of a new omega
//---------------------------------------
// t't
// t'r
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dt.dval, dt.dval, d1, d2, dskp.dval, queue ));
magma_zgetvector( 2, dskp.dval, 1, hskp.val, 1, queue );
// |t|
nrmt = magma_dsqrt( MAGMA_Z_REAL(hskp.val[0]) );
// rho = abs((t' * r) / (|t| * |r|))
rho = MAGMA_D_ABS( MAGMA_Z_REAL(hskp.val[1]) / (nrmt * nrmr) );
// om = (t' * r) / (|t| * |t|)
om = hskp.val[1] / hskp.val[0];
if ( rho < angle ) {
om = (om * angle) / rho;
}
//---------------------------------------
if ( MAGMA_Z_EQUAL(om, MAGMA_Z_ZERO) ) {
info = MAGMA_DIVERGENCE;
break;
}
// update approximation vector
// x = x + om * v
magma_zaxpy( x->num_rows, om, dv.dval, 1, x->dval, 1, queue );
// update residual vector
// r = r - om * t
magma_zaxpy( dr.num_rows, -om, dt.dval, 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// residual norm
nrmr = magma_dznrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
magma_zidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queue );
// t't
// t'rs
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queue ));
magma_zgetvector( 2, &dskp.dval[2], 1, &hskp.val[2], 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = hskp.val[3] / hskp.val[2];
// rs = rs - gamma * (rs - r)
magma_zaxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_zidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queue );
// |rs|
nrmr = magma_dznrm2( drs.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
break;
}
}
while ( solver_par->numiter + 1 <= solver_par->maxiter );
// smoothing enabled
if ( smoothing > 0 ) {
// x = xs
magma_zcopyvector( x->num_rows, dxs.dval, 1, x->dval, 1, queue );
// r = rs
magma_zcopyvector( dr.num_rows, drs.dval, 1, dr.dval, 1, queue );
}
// get last iteration timing
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t)tempo2 - tempo1;
//--------------STOP TIME----------------
// get final stats
solver_par->iter_res = nrmr;
CHECK( magma_zresidualvec( A, b, *x, &dr, &residual, queue ));
solver_par->final_res = residual;
// set solver conclusion
if ( info != MAGMA_SUCCESS && info != MAGMA_DIVERGENCE ) {
if ( solver_par->init_res > solver_par->final_res ) {
info = MAGMA_SLOW_CONVERGENCE;
}
}
cleanup:
// free resources
// smoothing enabled
if ( smoothing > 0 ) {
drs.dval = NULL; // needed because its pointer is redirected to dtt
magma_zmfree( &dxs, queue );
magma_zmfree( &drs, queue );
magma_zmfree( &dtt, queue );
}
dr.dval = NULL; // needed because its pointer is redirected to dt
dGcol.dval = NULL; // needed because its pointer is redirected to dG
magma_zmfree( &dr, queue );
magma_zmfree( &dP, queue );
magma_zmfree( &dP1, queue );
magma_zmfree( &dG, queue );
magma_zmfree( &dGcol, queue );
magma_zmfree( &dU, queue );
magma_zmfree( &dM, queue );
magma_zmfree( &hMdiag, queue );
magma_zmfree( &df, queue );
magma_zmfree( &dt, queue );
magma_zmfree( &dc, queue );
magma_zmfree( &dv, queue );
magma_zmfree( &dlu, queue );
magma_zmfree( &dskp, queue );
magma_zmfree( &dalpha, queue );
magma_zmfree( &dbeta, queue );
magma_zmfree( &hskp, queue );
magma_zmfree( &halpha, queue );
magma_zmfree( &hbeta, queue );
magma_free( d1 );
magma_free( d2 );
solver_par->info = info;
return info;
/* magma_zpidr_merge */
}
/**
Purpose
-------
ZLAQPS computes a step of QR factorization with column pivoting
of a complex M-by-N matrix A by using Blas-3. It tries to factorize
NB columns from A starting from the row OFFSET+1, and updates all
of the matrix with Blas-3 xGEMM.
In some cases, due to catastrophic cancellations, it cannot
factorize NB columns. Hence, the actual number of factorized
columns is returned in KB.
Block A(1:OFFSET,1:N) is accordingly pivoted, but not factorized.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0
@param[in]
offset INTEGER
The number of rows of A that have been factorized in
previous steps.
@param[in]
nb INTEGER
The number of columns to factorize.
@param[out]
kb INTEGER
The number of columns actually factorized.
@param[in,out]
A COMPLEX_16 array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, block A(OFFSET+1:M,1:KB) is the triangular
factor obtained and block A(1:OFFSET,1:N) has been
accordingly pivoted, but no factorized.
The rest of the matrix, block A(OFFSET+1:M,KB+1:N) has
been updated.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[in,out]
jpvt INTEGER array, dimension (N)
JPVT(I) = K <==> Column K of the full matrix A has been
permuted into position I in AP.
@param[out]
tau COMPLEX_16 array, dimension (KB)
The scalar factors of the elementary reflectors.
@param[in,out]
vn1 DOUBLE PRECISION array, dimension (N)
The vector with the partial column norms.
@param[in,out]
vn2 DOUBLE PRECISION array, dimension (N)
The vector with the exact column norms.
@param[in,out]
auxv COMPLEX_16 array, dimension (NB)
Auxiliar vector.
@param[in,out]
F COMPLEX_16 array, dimension (LDF,NB)
Matrix F' = L*Y'*A.
@param[in]
ldf INTEGER
The leading dimension of the array F. LDF >= max(1,N).
@ingroup magma_zgeqp3_aux
********************************************************************/
extern "C" magma_int_t
magma_zlaqps(
magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex_ptr dA, magma_int_t ldda,
magma_int_t *jpvt, magmaDoubleComplex *tau, double *vn1, double *vn2,
magmaDoubleComplex *auxv,
magmaDoubleComplex *F, magma_int_t ldf,
magmaDoubleComplex_ptr dF, magma_int_t lddf)
{
#define A(i, j) (A + (i) + (j)*(lda ))
#define dA(i, j) (dA + (i) + (j)*(ldda))
#define F(i, j) (F + (i) + (j)*(ldf ))
#define dF(i, j) (dF + (i) + (j)*(lddf))
magmaDoubleComplex c_zero = MAGMA_Z_MAKE( 0.,0.);
magmaDoubleComplex c_one = MAGMA_Z_MAKE( 1.,0.);
magmaDoubleComplex c_neg_one = MAGMA_Z_MAKE(-1.,0.);
magma_int_t ione = 1;
magma_int_t i__1, i__2;
double d__1;
magmaDoubleComplex z__1;
magma_int_t j, k, rk;
magmaDoubleComplex Akk;
magma_int_t pvt;
double temp, temp2, tol3z;
magma_int_t itemp;
magma_int_t lsticc;
magma_int_t lastrk;
lastrk = min( m, n + offset );
tol3z = magma_dsqrt( lapackf77_dlamch("Epsilon"));
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
lsticc = 0;
k = 0;
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// subtract 1 from Fortran idamax; pvt, k are 0-based.
i__1 = n-k;
pvt = k + blasf77_idamax( &i__1, &vn1[k], &ione ) - 1;
if (pvt != k) {
if (pvt >= nb) {
/* 1. Start copy from GPU */
magma_zgetmatrix_async( m - offset - nb, 1,
dA(offset + nb, pvt), ldda,
A (offset + nb, pvt), lda, queue );
}
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
blasf77_zswap( &i__1, F(pvt,0), &ldf, F(k,0), &ldf );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
vn1[pvt] = vn1[k];
vn2[pvt] = vn2[k];
if (pvt < nb) {
/* no need of transfer if pivot is within the panel */
blasf77_zswap( &m, A(0, pvt), &ione, A(0, k), &ione );
}
else {
/* 1. Finish copy from GPU */
magma_queue_sync( queue );
/* 2. Swap as usual on CPU */
blasf77_zswap(&m, A(0, pvt), &ione, A(0, k), &ione);
/* 3. Restore the GPU */
magma_zsetmatrix_async( m - offset - nb, 1,
A (offset + nb, pvt), lda,
dA(offset + nb, pvt), ldda, queue );
}
}
/* Apply previous Householder reflectors to column K:
A(RK:M,K) := A(RK:M,K) - A(RK:M,1:K-1)*F(K,1:K-1)'.
Optimization: multiply with beta=0; wait for vector and subtract */
if (k > 0) {
#ifdef COMPLEX
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_Z_CONJ( *F(k,j) );
}
#endif
i__1 = m - rk;
i__2 = k;
blasf77_zgemv( MagmaNoTransStr, &i__1, &i__2,
&c_neg_one, A(rk, 0), &lda,
F(k, 0), &ldf,
&c_one, A(rk, k), &ione );
#ifdef COMPLEX
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_Z_CONJ( *F(k,j) );
}
#endif
}
/* Generate elementary reflector H(k). */
if (rk < m-1) {
i__1 = m - rk;
lapackf77_zlarfg( &i__1, A(rk, k), A(rk + 1, k), &ione, &tau[k] );
} else {
lapackf77_zlarfg( &ione, A(rk, k), A(rk, k), &ione, &tau[k] );
}
Akk = *A(rk, k);
*A(rk, k) = c_one;
/* Compute Kth column of F:
Compute F(K+1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) on the GPU */
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Send the vector to the GPU */
magma_zsetmatrix( i__1, 1, A(rk, k), lda, dA(rk,k), ldda, queue );
/* Multiply on GPU */
// was CALL ZGEMV( 'Conjugate transpose', M-RK+1, N-K,
// TAU( K ), A( RK, K+1 ), LDA,
// A( RK, K ), 1,
// CZERO, F( K+1, K ), 1 )
magma_int_t i__3 = nb-k-1;
magma_int_t i__4 = i__2 - i__3;
magma_int_t i__5 = nb-k;
magma_zgemv( MagmaConjTrans, i__1 - i__5, i__2 - i__3,
tau[k], dA(rk +i__5, k+1+i__3), ldda,
dA(rk +i__5, k ), ione,
c_zero, dF(k+1+i__3, k ), ione, queue );
magma_zgetmatrix_async( i__2-i__3, 1,
dF(k + 1 +i__3, k), i__2,
F (k + 1 +i__3, k), i__2, queue );
blasf77_zgemv( MagmaConjTransStr, &i__1, &i__3,
&tau[k], A(rk, k+1), &lda,
A(rk, k ), &ione,
&c_zero, F(k+1, k ), &ione );
magma_queue_sync( queue );
blasf77_zgemv( MagmaConjTransStr, &i__5, &i__4,
&tau[k], A(rk, k+1+i__3), &lda,
A(rk, k ), &ione,
&c_one, F(k+1+i__3, k ), &ione );
}
/* Padding F(1:K,K) with zeros. */
for (j = 0; j < k; ++j) {
*F(j, k) = c_zero;
}
/* Incremental updating of F:
F(1:N,K) := F(1:N,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K). */
if (k > 0) {
i__1 = m - rk;
i__2 = k;
z__1 = MAGMA_Z_NEGATE( tau[k] );
blasf77_zgemv( MagmaConjTransStr, &i__1, &i__2,
&z__1, A(rk, 0), &lda,
A(rk, k), &ione,
&c_zero, auxv, &ione );
i__1 = k;
blasf77_zgemv( MagmaNoTransStr, &n, &i__1,
&c_one, F(0,0), &ldf,
auxv, &ione,
&c_one, F(0,k), &ione );
}
/* Optimization: On the last iteration start sending F back to the GPU */
/* Update the current row of A:
A(RK,K+1:N) := A(RK,K+1:N) - A(RK,1:K)*F(K+1:N,1:K)'. */
if (k < n-1) {
i__1 = n - k - 1;
i__2 = k + 1;
blasf77_zgemm( MagmaNoTransStr, MagmaConjTransStr, &ione, &i__1, &i__2,
&c_neg_one, A(rk, 0 ), &lda,
F(k+1,0 ), &ldf,
&c_one, A(rk, k+1), &lda );
}
/* Update partial column norms. */
if (rk < lastrk) {
for (j = k + 1; j < n; ++j) {
if (vn1[j] != 0.) {
/* NOTE: The following 4 lines follow from the analysis in
Lapack Working Note 176. */
temp = MAGMA_Z_ABS( *A(rk,j) ) / vn1[j];
temp = max( 0., ((1. + temp) * (1. - temp)) );
d__1 = vn1[j] / vn2[j];
temp2 = temp * (d__1 * d__1);
if (temp2 <= tol3z) {
vn2[j] = (double) lsticc;
lsticc = j;
} else {
vn1[j] *= magma_dsqrt(temp);
}
}
}
}
*A(rk, k) = Akk;
++k;
}
// leave k as the last column done
--k;
*kb = k + 1;
rk = offset + *kb - 1;
/* Apply the block reflector to the rest of the matrix:
A(OFFSET+KB+1:M,KB+1:N) := A(OFFSET+KB+1:M,KB+1:N) - A(OFFSET+KB+1:M,1:KB)*F(KB+1:N,1:KB)' */
if (*kb < min(n, m - offset)) {
i__1 = m - rk - 1;
i__2 = n - *kb;
/* Send F to the GPU */
magma_zsetmatrix( i__2, *kb,
F (*kb, 0), ldf,
dF(*kb, 0), i__2, queue );
magma_zgemm( MagmaNoTrans, MagmaConjTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), i__2,
c_one, dA(rk+1, *kb), ldda, queue );
}
/* Recomputation of difficult columns. */
while( lsticc > 0 ) {
itemp = (magma_int_t)(vn2[lsticc] >= 0. ? floor(vn2[lsticc] + .5) : -floor(.5 - vn2[lsticc]));
i__1 = m - rk - 1;
if (lsticc <= nb) {
vn1[lsticc] = magma_cblas_dznrm2( i__1, A(rk+1,lsticc), ione );
}
else {
/* Where is the data, CPU or GPU ? */
double r1, r2;
r1 = magma_cblas_dznrm2( nb-k, A(rk+1,lsticc), ione );
r2 = magma_dznrm2( m-offset-nb, dA(offset + nb + 1, lsticc), ione, queue );
//vn1[lsticc] = magma_dznrm2( i__1, dA(rk + 1, lsticc), ione, queue );
vn1[lsticc] = magma_dsqrt(r1*r1 + r2*r2);
}
/* NOTE: The computation of VN1( LSTICC ) relies on the fact that
SNRM2 does not fail on vectors with norm below the value of SQRT(DLAMCH('S')) */
vn2[lsticc] = vn1[lsticc];
lsticc = itemp;
}
magma_queue_destroy( queue );
return MAGMA_SUCCESS;
} /* magma_zlaqps */
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 */
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 */
/* ////////////////////////////////////////////////////////////////////////////
-- testing any solver
*/
int main( int argc, char** argv )
{
magma_int_t info = 0;
TESTING_INIT();
magma_queue_t queue=NULL;
magma_queue_create( 0, &queue );
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magma_z_matrix A={Magma_CSR}, B_d={Magma_CSR};
magma_z_matrix x={Magma_CSR}, b={Magma_CSR};
int i=1;
while( i < argc ) {
if ( strcmp("LAPLACE2D", argv[i]) == 0 && i+1 < argc ) { // Laplace test
i++;
magma_int_t laplace_size = atoi( argv[i] );
CHECK( magma_zm_5stencil( laplace_size, &A, queue ));
} else { // file-matrix test
CHECK( magma_z_csr_mtx( &A, argv[i], queue ));
}
printf( "\n# matrix info: %d-by-%d with %d nonzeros\n\n",
int(A.num_rows), int(A.num_cols), int(A.nnz) );
magma_int_t n = A.num_rows;
CHECK( magma_zmtransfer( A, &B_d, Magma_CPU, Magma_DEV, queue ));
// vectors and initial guess
CHECK( magma_zvinit( &b, Magma_DEV, A.num_cols, 1, zero, queue ));
CHECK( magma_zvinit( &x, Magma_DEV, A.num_cols, 1, one, queue ));
CHECK( magma_zprint_vector( b, 90, 10, queue ));
CHECK( magma_zprint_matrix( A, queue ));
printf("\n\n\n");
CHECK( magma_zprint_matrix( B_d, queue ));
double res;
res = magma_dznrm2(n, b.dval, 1, queue );
printf("norm0: %f\n", res);
CHECK( magma_z_spmv( one, B_d, x, zero, b, queue )); // b = A x
CHECK( magma_zprint_vector( b, 0, 100, queue ));
CHECK( magma_zprint_vector( b, b.num_rows-10, 10, queue ));
res = magma_dznrm2( n, b.dval, 1, queue );
printf("norm: %f\n", res);
CHECK( magma_zresidual( B_d, x, b, &res, queue));
printf("res: %f\n", res);
magma_zmfree(&B_d, queue );
magma_zmfree(&A, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
i++;
}
cleanup:
magma_zmfree(&A, queue );
magma_zmfree(&B_d, queue );
magma_zmfree(&x, queue );
magma_zmfree(&b, queue );
magma_queue_destroy( queue );
magma_finalize();
return info;
}
extern "C" magma_int_t
magma_zbicgstab_merge3(
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_BICGSTABMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// solver variables
magmaDoubleComplex alpha, beta, omega, rho_old, rho_new, *skp_h={0};
double nom, nom0, betanom, nomb;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows;
// workspace
magma_z_matrix q={Magma_CSR}, r={Magma_CSR}, rr={Magma_CSR}, p={Magma_CSR}, v={Magma_CSR}, s={Magma_CSR}, t={Magma_CSR};
magmaDoubleComplex *d1=NULL, *d2=NULL, *skp=NULL;
d1 = NULL;
d2 = NULL;
skp = NULL;
CHECK( magma_zmalloc( &d1, dofs*(2) ));
CHECK( magma_zmalloc( &d2, dofs*(2) ));
// array for the parameters
CHECK( magma_zmalloc( &skp, 8 ));
// skp = [alpha|beta|omega|rho_old|rho|nom|tmp1|tmp2]
CHECK( magma_zvinit( &q, Magma_DEV, dofs*6, 1, c_zero, queue ));
// q = rr|r|p|v|s|t
rr.memory_location = Magma_DEV; rr.dval = NULL; rr.num_rows = rr.nnz = dofs; rr.num_cols = 1; rr.storage_type = Magma_DENSE;
r.memory_location = Magma_DEV; r.dval = NULL; r.num_rows = r.nnz = dofs; r.num_cols = 1; r.storage_type = Magma_DENSE;
p.memory_location = Magma_DEV; p.dval = NULL; p.num_rows = p.nnz = dofs; p.num_cols = 1; p.storage_type = Magma_DENSE;
v.memory_location = Magma_DEV; v.dval = NULL; v.num_rows = v.nnz = dofs; v.num_cols = 1; v.storage_type = Magma_DENSE;
s.memory_location = Magma_DEV; s.dval = NULL; s.num_rows = s.nnz = dofs; s.num_cols = 1; s.storage_type = Magma_DENSE;
t.memory_location = Magma_DEV; t.dval = NULL; t.num_rows = t.nnz = dofs; t.num_cols = 1; t.storage_type = Magma_DENSE;
rr.dval = q(0);
r.dval = q(1);
p.dval = q(2);
v.dval = q(3);
s.dval = q(4);
t.dval = q(5);
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
magma_zcopy( dofs, r.dval, 1, q(0), 1, queue ); // rr = r
magma_zcopy( dofs, r.dval, 1, q(1), 1, queue ); // q = 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. );
beta = rho_new;
solver_par->init_res = nom0;
// array on host for the parameters
CHECK( magma_zmalloc_cpu( &skp_h, 8 ));
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
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;
}
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_Z_MAKE(nom, 0.0);
magma_zsetvector( 8, skp_h, 1, skp, 1, queue );
CHECK( magma_z_spmv( c_one, A, r, c_zero, v, queue )); // z = A r
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if( nom0 < solver_par->atol ||
nom0/nomb < solver_par->rtol ){
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++;
// computes p=r+beta*(p-omega*v)
CHECK( magma_zbicgmerge1( dofs, skp, v.dval, r.dval, p.dval, queue ));
CHECK( magma_z_spmv( c_one, A, p, c_zero, v, queue )); // v = Ap
solver_par->spmv_count++;
CHECK( magma_zmdotc( dofs, 1, q.dval, v.dval, d1, d2, skp, queue ));
CHECK( magma_zbicgmerge4( 1, skp, queue ));
CHECK( magma_zbicgmerge2( dofs, skp, r.dval, v.dval, s.dval, queue )); // s=r-alpha*v
CHECK( magma_z_spmv( c_one, A, s, c_zero, t, queue )); // t=As
solver_par->spmv_count++;
CHECK( magma_zmdotc( dofs, 2, q.dval+4*dofs, t.dval, d1, d2, skp+6, queue ));
CHECK( magma_zbicgmerge4( 2, skp, queue ));
CHECK( magma_zbicgmerge_xrbeta( dofs, d1, d2, q.dval, r.dval, p.dval,
s.dval, t.dval, x->dval, skp, queue ));
// check stopping criterion
magma_zgetvector_async( 1 , skp+5, 1, skp_h+5, 1, queue );
betanom = sqrt(MAGMA_Z_REAL(skp_h[5]));
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) betanom;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( betanom < solver_par->atol ||
betanom/nomb < solver_par->rtol ) {
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 = betanom;
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) 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->atol ||
solver_par->iter_res/solver_par->init_res < solver_par->rtol ){
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(&q, queue ); // frees all vectors
magma_free(d1);
magma_free(d2);
magma_free( skp );
magma_free_cpu( skp_h );
solver_par->info = info;
return info;
} /* zbicgstab_merge */
extern "C" magma_int_t
magma_ztfqmr_unrolled(
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_TFQMR;
solver_par->numiter = 0;
solver_par->spmv_count = 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, 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},
d={Magma_CSR}, w={Magma_CSR}, v={Magma_CSR},
u_mp1={Magma_CSR}, u_m={Magma_CSR}, Au={Magma_CSR},
Ad={Magma_CSR}, Au_new={Magma_CSR};
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( &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_mp1.dval, 1, queue );
CHECK( magma_z_spmv( c_one, A, u_mp1, 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_zaxpy( dofs, -alpha, v.dval, 1, u_mp1.dval, 1, queue ); // u_mp1 = u_mp_1 - alpha*v;
magma_zaxpy( dofs, -alpha, Au.dval, 1, w.dval, 1, queue ); // w = w - alpha*Au;
magma_zscal( dofs, sigma, d.dval, 1, queue );
magma_zaxpy( dofs, c_one, u_mp1.dval, 1, d.dval, 1, queue ); // d = u_mp1 + sigma*d;
//magma_zscal( dofs, sigma, Ad.dval, 1, queue );
//magma_zaxpy( dofs, c_one, Au.dval, 1, Ad.dval, 1, queue ); // Ad = Au + sigma*Ad;
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;
printf("sigma: %f+%fi\n", MAGMA_Z_REAL(sigma), MAGMA_Z_IMAG(sigma) );
CHECK( magma_z_spmv( c_one, A, d, c_zero, Ad, queue )); // Au_new = A u_mp1
solver_par->spmv_count++;
magma_zaxpy( dofs, eta, d.dval, 1, x->dval, 1, queue ); // x = x + eta * d
magma_zaxpy( dofs, -eta, Ad.dval, 1, r.dval, 1, queue ); // r = r - eta * Ad
// here starts the second part of the loop #################################
magma_zaxpy( dofs, -alpha, Au.dval, 1, w.dval, 1, queue ); // w = w - alpha*Au;
magma_zscal( dofs, sigma, d.dval, 1, queue );
magma_zaxpy( dofs, c_one, u_mp1.dval, 1, d.dval, 1, queue ); // d = u_mp1 + sigma*d;
magma_zscal( dofs, sigma, Ad.dval, 1, queue );
magma_zaxpy( dofs, c_one, Au.dval, 1, Ad.dval, 1, queue ); // Ad = Au + sigma*Ad;
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_zaxpy( dofs, eta, d.dval, 1, x->dval, 1, queue ); // x = x + eta * d
magma_zaxpy( dofs, -eta, Ad.dval, 1, r.dval, 1, queue ); // r = r - eta * Ad
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;
}
// do this every loop as unrolled
rho_l = rho;
rho = magma_zdotc( dofs, w.dval, 1, r_tld.dval, 1, queue );
beta = rho / rho_l;
magma_zscal( dofs, beta, u_mp1.dval, 1, queue );
magma_zaxpy( dofs, c_one, w.dval, 1, u_mp1.dval, 1, queue ); // u_mp1 = w + beta*u_mp1;
CHECK( magma_z_spmv( c_one, A, u_mp1, c_zero, Au_new, queue )); // Au_new = A u_mp1
solver_par->spmv_count++;
// do this every loop as unrolled
magma_zscal( dofs, beta*beta, v.dval, 1, queue );
magma_zaxpy( dofs, beta, Au.dval, 1, v.dval, 1, queue );
magma_zaxpy( dofs, c_one, Au_new.dval, 1, v.dval, 1, queue ); // v = Au_new + beta*(Au+beta*v);
magma_zcopy( dofs, Au_new.dval, 1, Au.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(&d, queue );
magma_zmfree(&Au, queue );
magma_zmfree(&Au_new, queue );
magma_zmfree(&Ad, queue );
solver_par->info = info;
return info;
} /* magma_zfqmr_unrolled */
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 */
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_zcg_res(
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_CG;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// solver variables
magmaDoubleComplex alpha, beta;
double nom, nom0, r0, res, nomb;
magmaDoubleComplex den, gammanew, gammaold = MAGMA_Z_MAKE(1.0,0.0);
// local variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows* b.num_cols;
// GPU workspace
magma_z_matrix r={Magma_CSR}, p={Magma_CSR}, q={Magma_CSR};
CHECK( magma_zvinit( &r, 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( &q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
magma_zcopy( dofs, r.dval, 1, p.dval, 1, queue ); // p = h
nom = MAGMA_Z_ABS( magma_zdotc( dofs, r.dval, 1, r.dval, 1, queue) );
CHECK( magma_z_spmv( c_one, A, p, c_zero, q, queue )); // q = A p
solver_par->spmv_count++;
den = magma_zdotc( dofs, p.dval, 1, q.dval, 1, queue ); // den = p dot q
solver_par->init_res = nom0;
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 ( nom < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// check positive definite
if ( MAGMA_Z_ABS(den) <= 0.0 ) {
info = MAGMA_NONSPD;
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++;
gammanew = magma_zdotc( dofs, r.dval, 1, r.dval, 1, queue );
// gn = < r,r>
if ( solver_par->numiter == 1 ) {
magma_zcopy( dofs, r.dval, 1, p.dval, 1, queue ); // p = r
} else {
beta = (gammanew/gammaold); // beta = gn/go
magma_zscal( dofs, beta, p.dval, 1, queue ); // p = beta*p
magma_zaxpy( dofs, c_one, r.dval, 1, p.dval, 1, queue ); // p = p + r
}
CHECK( magma_z_spmv( c_one, A, p, c_zero, q, queue )); // q = A p
solver_par->spmv_count++;
den = magma_zdotc( dofs, p.dval, 1, q.dval, 1, queue );
// den = p dot q
alpha = gammanew / den;
magma_zaxpy( dofs, alpha, p.dval, 1, x->dval, 1, queue ); // x = x + alpha p
magma_zaxpy( dofs, -alpha, q.dval, 1, r.dval, 1, queue ); // r = r - alpha q
gammaold = gammanew;
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(&p, queue );
magma_zmfree(&q, queue );
solver_par->info = info;
return info;
} /* magma_zcg */
extern "C" magma_int_t
magma_zpcg_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_PCGMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// solver variables
magmaDoubleComplex alpha, beta, gamma, rho, tmp1, *skp_h={0};
double nom, nom0, r0, res, nomb;
magmaDoubleComplex den;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows*b.num_cols;
magma_z_matrix r={Magma_CSR}, d={Magma_CSR}, z={Magma_CSR}, h={Magma_CSR},
rt={Magma_CSR};
magmaDoubleComplex *d1=NULL, *d2=NULL, *skp=NULL;
// GPU workspace
CHECK( magma_zvinit( &r, 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( &z, 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( &h, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zmalloc( &d1, dofs*(2) ));
CHECK( magma_zmalloc( &d2, dofs*(2) ));
// array for the parameters
CHECK( magma_zmalloc( &skp, 7 ));
// skp = [alpha|beta|gamma|rho|tmp1|tmp2|res]
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
// preconditioner
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, r, &rt, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, rt, &h, precond_par, queue ));
magma_zcopy( dofs, h.dval, 1, d.dval, 1, queue );
nom = MAGMA_Z_ABS( magma_zdotc( dofs, r.dval, 1, h.dval, 1, queue ));
CHECK( magma_z_spmv( c_one, A, d, c_zero, z, queue )); // z = A d
den = magma_zdotc( dofs, d.dval, 1, z.dval, 1, queue ); // den = d'* z
solver_par->init_res = nom0;
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 ( nom < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// check positive definite
if ( MAGMA_Z_ABS(den) <= 0.0 ) {
info = MAGMA_NONSPD;
goto cleanup;
}
// array on host for the parameters
CHECK( magma_zmalloc_cpu( &skp_h, 7 ));
alpha = rho = gamma = tmp1 = c_one;
beta = magma_zdotc( dofs, h.dval, 1, r.dval, 1, queue );
skp_h[0]=alpha;
skp_h[1]=beta;
skp_h[2]=gamma;
skp_h[3]=rho;
skp_h[4]=tmp1;
skp_h[5]=MAGMA_Z_MAKE(nom, 0.0);
skp_h[6]=MAGMA_Z_MAKE(nom, 0.0);
magma_zsetvector( 7, skp_h, 1, skp, 1, queue );
//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++;
// computes SpMV and dot product
CHECK( magma_zcgmerge_spmv1( A, d1, d2, d.dval, z.dval, skp, queue ));
solver_par->spmv_count++;
if( precond_par->solver == Magma_JACOBI ){
CHECK( magma_zjcgmerge_xrbeta( dofs, d1, d2, precond_par->d.dval, x->dval, r.dval, d.dval, z.dval, h.dval, skp, queue ));
}
else if( precond_par->solver == Magma_NONE ){
// updates x, r
CHECK( magma_zpcgmerge_xrbeta1( dofs, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// computes scalars and updates d
CHECK( magma_zpcgmerge_xrbeta2( dofs, d1, d2, r.dval, r.dval, d.dval, skp, queue ));
} else {
// updates x, r
CHECK( magma_zpcgmerge_xrbeta1( dofs, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// preconditioner in between
tempop1 = magma_sync_wtime( queue );
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, r, &rt, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, rt, &h, precond_par, queue ));
// magma_zcopy( dofs, r.dval, 1, h.dval, 1 );
tempop2 = magma_sync_wtime( queue );
precond_par->runtime += tempop2-tempop1;
// computes scalars and updates d
CHECK( magma_zpcgmerge_xrbeta2( dofs, d1, d2, h.dval, r.dval, d.dval, skp, queue ));
}
//if( solver_par->numiter==1){
// magma_zcopy( dofs, h.dval, 1, d.dval, 1 );
//}
// updates x, r, computes scalars and updates d
//CHECK( magma_zcgmerge_xrbeta( dofs, d1, d2, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// check stopping criterion (asynchronous copy)
magma_zgetvector( 1 , skp+6, 1, skp_h+6, 1, queue );
res = sqrt(MAGMA_Z_ABS(skp_h[6]));
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->atol ||
solver_par->iter_res/solver_par->init_res < solver_par->rtol ){
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;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&z, queue );
magma_zmfree(&d, queue );
magma_zmfree(&rt, queue );
magma_zmfree(&h, queue );
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
magma_free_cpu( skp_h );
solver_par->info = info;
return info;
} /* magma_zpcg_merge */
extern "C" magma_int_t
magma_zbombard_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;
// queue variables
const magma_queue_t squeue = 0; // synchronous kernel queues
const magma_int_t nqueues = 3; // number of queues
magma_queue_t queues[nqueues];
magma_int_t q1flag = 0;
// set asynchronous kernel queues
//printf("%% Kernel queues: (orig, queue) = (%p, %p)\n", (void *)orig_queue, (void *)queue);
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
if ( queue != squeue ) {
queues[1] = queue;
q1flag = 0;
} else {
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &(queues[1]) );
queue = queues[1];
q1flag = 1;
}
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &(queues[2]) );
for (int i = 0; i < nqueues; ++i ) {
; //printf("Kernel queue #%d = %p\n", i, (void *)queues[i]);
}
// 1=QMR, 2=CGS, 3+BiCGSTAB
magma_int_t flag = 0;
int mdot = 1;
// prepare solver feedback
solver_par->solver = Magma_BOMBARD;
solver_par->numiter = 0;
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
// solver variables
double nom0, r0, res, Q_res, C_res, B_res, nomb;
//QMR
magmaDoubleComplex Q_rho = c_one, Q_rho1 = c_one, Q_eta = -c_one , Q_pds = c_one,
Q_thet = c_one, Q_thet1 = c_one, Q_epsilon = c_one,
Q_beta = c_one, Q_delta = c_one, Q_pde = c_one, Q_rde = c_one,
Q_gamm = c_one, Q_gamm1 = c_one, Q_psi = c_one;
//CGS
magmaDoubleComplex C_rho, C_rho_l = c_one, C_alpha, C_beta;
//BiCGSTAB
magmaDoubleComplex B_alpha, B_beta, B_omega, B_rho_old, B_rho_new;
magma_int_t dofs = A.num_rows* b.num_cols;
// need to transpose the matrix
// GPU workspace
// overall initial residual
magma_z_matrix r_tld={Magma_CSR};
// QMR
magma_z_matrix AT = {Magma_CSR}, Q_r={Magma_CSR}, Q_x={Magma_CSR},
Q_v={Magma_CSR}, Q_w={Magma_CSR}, Q_wt={Magma_CSR},
Q_d={Magma_CSR}, Q_s={Magma_CSR}, Q_z={Magma_CSR}, Q_q={Magma_CSR},
Q_p={Magma_CSR}, Q_pt={Magma_CSR}, Q_y={Magma_CSR};
// CGS
magma_z_matrix C_r={Magma_CSR}, C_rt={Magma_CSR}, C_x={Magma_CSR},
C_p={Magma_CSR}, C_q={Magma_CSR}, C_u={Magma_CSR}, C_v={Magma_CSR}, C_t={Magma_CSR},
C_p_hat={Magma_CSR}, C_q_hat={Magma_CSR}, C_u_hat={Magma_CSR}, C_v_hat={Magma_CSR};
//BiCGSTAB
magma_z_matrix B_r={Magma_CSR}, B_x={Magma_CSR}, B_p={Magma_CSR}, B_v={Magma_CSR},
B_s={Magma_CSR}, B_t={Magma_CSR};
// multi-vector for block-SpMV
magma_z_matrix SpMV_in_1={Magma_CSR}, SpMV_out_1={Magma_CSR},
SpMV_in_2={Magma_CSR}, SpMV_out_2={Magma_CSR};
// workspace for zmzdotc
magma_z_matrix d1={Magma_CSR}, d2={Magma_CSR}, skp={Magma_CSR};
CHECK( magma_zvinit( &r_tld, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &SpMV_in_1, Magma_DEV, A.num_rows, b.num_cols * 3, c_zero, queue ));
CHECK( magma_zvinit( &SpMV_out_1, Magma_DEV, A.num_rows, b.num_cols * 3, c_zero, queue ));
CHECK( magma_zvinit( &SpMV_in_2, Magma_DEV, A.num_rows, b.num_cols * 3, c_zero, queue ));
CHECK( magma_zvinit( &SpMV_out_2, Magma_DEV, A.num_rows, b.num_cols * 3, c_zero, queue ));
CHECK( magma_zvinit( &d1, Magma_DEV, A.num_rows, b.num_cols * 4, c_zero, queue ));
CHECK( magma_zvinit( &d2, Magma_DEV, A.num_rows, b.num_cols * 4, c_zero, queue ));
CHECK( magma_zvinit( &skp, Magma_CPU, 4, 1, c_zero, queue ));
// QMR
CHECK( magma_zvinit( &Q_r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_w, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_wt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_d, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_s, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_pt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_y, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &Q_x, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// QMR
CHECK( magma_zvinit( &C_r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_rt,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_x,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_p_hat, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_q, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_q_hat, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_u, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_u_hat, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_v_hat, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &C_t, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// BiCGSTAB
CHECK( magma_zvinit( &B_r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &B_x,Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &B_p, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &B_v, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &B_s, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &B_t, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r_tld, &nom0, queue));
solver_par->init_res = nom0;
res = nom0;
// QMR
magma_zcopy( dofs, r_tld.dval, 1, Q_r.dval, 1, queue );
magma_zcopy( dofs, r_tld.dval, 1, Q_y.dval, 1, queue );
magma_zcopy( dofs, r_tld.dval, 1, Q_v.dval, 1, queue );
magma_zcopy( dofs, r_tld.dval, 1, Q_wt.dval, 1, queue );
magma_zcopy( dofs, r_tld.dval, 1, Q_z.dval, 1, queue );
magma_zcopy( dofs, x->dval, 1, Q_x.dval, 1, queue );
// transpose the matrix
magma_zmtransposeconjugate( A, &AT, queue );
// CGS
magma_zcopy( dofs, r_tld.dval, 1, C_r.dval, 1, queue );
magma_zcopy( dofs, x->dval, 1, C_x.dval, 1, queue );
// BiCGSTAB
magma_zcopy( dofs, r_tld.dval, 1, B_r.dval, 1, queue );
magma_zcopy( dofs, x->dval, 1, B_x.dval, 1, queue );
CHECK( magma_z_spmv( c_one, A, B_r, c_zero, B_v, queue ));
magma_zcopy( dofs, B_v.dval, 1, SpMV_out_1.dval+2*dofs, 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;
}
Q_psi = magma_zsqrt( magma_zdotc( dofs, Q_z.dval, 1, Q_z.dval, 1), queue );
Q_rho = magma_zsqrt( magma_zdotc( dofs, Q_y.dval, 1, Q_y.dval, 1), queue );
// BiCGSTAB
B_rho_new = magma_zdotc( dofs, B_r.dval, 1, B_r.dval, 1, queue );
B_rho_old = B_omega = B_alpha = MAGMA_Z_MAKE( 1.0, 0. );
// v = y / rho
// y = y / rho
// w = wt / psi
// z = z / psi
magma_zqmr_1(
b.num_rows,
b.num_cols,
Q_rho,
Q_psi,
Q_y.dval,
Q_z.dval,
Q_v.dval,
Q_w.dval,
queue );
//Chronometry
real_Double_t tempo1, tempo2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
// start iteration
do
{
solver_par->numiter++;
if(mdot == 0){
//QMR: delta = z' * y;
Q_delta = magma_zdotc( dofs, Q_z.dval, 1, Q_y.dval, 1, queue );
//CGS: rho = r' * r_tld
C_rho = magma_zdotc( dofs, C_r.dval, 1, r_tld.dval, 1, queue );
// BiCGSTAB
B_rho_old = B_rho_new;
B_rho_new = magma_zdotc( dofs, r_tld.dval, 1, B_r.dval, 1, queue ); // rho=<rr,r>
B_beta = B_rho_new/B_rho_old * B_alpha/B_omega; // beta=rho/rho_old *alpha/omega
}else{
B_rho_old = B_rho_new;
magma_zmdotc4(
b.num_rows,
Q_z.dval,
Q_y.dval,
r_tld.dval,
C_r.dval,
r_tld.dval,
B_r.dval,
r_tld.dval,
B_r.dval,
d1.dval,
d2.dval,
skp.val,
queue );
Q_delta = skp.val[0];
C_rho = skp.val[1];
B_rho_new = skp.val[2];
B_beta = B_rho_new/B_rho_old * B_alpha/B_omega;
}
if( solver_par->numiter == 1 ){
//QMR: p = y;
//QMR: q = z;
magma_zcopy( dofs, Q_y.dval, 1, SpMV_in_1.dval, 1, queue );
magma_zcopy( dofs, Q_z.dval, 1, SpMV_in_2.dval, 1, queue );
//QMR: u = r;
//QMR: p = r;
magma_zcgs_2(
b.num_rows,
b.num_cols,
C_r.dval,
C_u.dval,
SpMV_in_1.dval+dofs,
queue );
}
else{
Q_pde = Q_psi * Q_delta / Q_epsilon;
Q_rde = Q_rho * MAGMA_Z_CNJG(Q_delta/Q_epsilon);
C_beta = C_rho / C_rho_l;
//QMR p = y - pde * p
//QMR q = z - rde * q
magma_zqmr_2(
b.num_rows,
b.num_cols,
Q_pde,
Q_rde,
Q_y.dval,
Q_z.dval,
SpMV_in_1.dval,
SpMV_in_2.dval,
queue );
//CGS: u = r + beta*q;
//CGS: p = u + beta*( q + beta*p );
magma_zcgs_1(
b.num_rows,
b.num_cols,
C_beta,
C_r.dval,
C_q.dval,
C_u.dval,
SpMV_in_1.dval+dofs,
queue );
}
// BiCGSTAB: p = r + beta * ( p - omega * v )
magma_zbicgstab_1(
b.num_rows,
b.num_cols,
B_beta,
B_omega,
B_r.dval,
SpMV_out_1.dval+2*dofs,
SpMV_in_1.dval+2*dofs,
queue );
/*
//QMR
CHECK( magma_z_spmv( c_one, A, Q_p, c_zero, Q_pt, queue ));
//CGS
CHECK( magma_z_spmv( c_one, A, C_p, c_zero, C_v_hat, queue ));
// BiCGSTAB
CHECK( magma_z_spmv( c_one, A, B_p, c_zero, B_v, queue ));
*/
// gather everything for a block-SpMV
//magma_zcopy( dofs, Q_p.dval, 1, SpMV_in_1.dval , 1, queue );
//magma_zcopy( dofs, C_p.dval, 1, SpMV_in_1.dval+dofs , 1, queue );
//magma_zcopy( dofs, B_p.dval, 1, SpMV_in_1.dval+2*dofs , 1, queue );
// block SpMV
CHECK( magma_z_spmv( c_one, A, SpMV_in_1, c_zero, SpMV_out_1, queue ));
// scatter results
//magma_zcopy( dofs, SpMV_out_1.dval , 1, Q_pt.dval, 1, queue );
//magma_zcopy( dofs, SpMV_out_1.dval+dofs , 1, C_v_hat.dval, 1, queue );
//magma_zcopy( dofs, SpMV_out_1.dval+2*dofs , 1, B_v.dval, 1, queue );
if( mdot == 0 ) {
//QMR: epsilon = q' * pt;
Q_epsilon = magma_zdotc( dofs, SpMV_in_2.dval, 1, SpMV_out_1.dval, 1, queue );
Q_beta = Q_epsilon / Q_delta;
//CGS: alpha = r_tld' * v_hat
C_alpha = C_rho / magma_zdotc( dofs, r_tld.dval, 1, SpMV_out_1.dval+dofs, 1, queue );
C_rho_l = C_rho;
//BiCGSTAB
B_alpha = B_rho_new / magma_zdotc( dofs, r_tld.dval, 1, SpMV_out_1.dval+2*dofs, 1, queue );
}else{
magma_zmdotc4(
b.num_rows,
SpMV_in_2.dval,
SpMV_out_1.dval,
r_tld.dval,
SpMV_out_1.dval+dofs,
r_tld.dval,
SpMV_out_1.dval+2*dofs,
r_tld.dval,
SpMV_out_1.dval+2*dofs,
d1.dval,
d2.dval,
skp.val,
queue );
Q_epsilon = skp.val[0];
Q_beta = Q_epsilon / Q_delta;
C_alpha = C_rho / skp.val[1];
C_rho_l = C_rho;
B_alpha = B_rho_new / skp.val[2];
}
//QMR: v = pt - beta * v
//QMR: y = v
magma_zqmr_3(
b.num_rows,
b.num_cols,
Q_beta,
SpMV_out_1.dval,
Q_v.dval,
Q_y.dval,
queue );
//CGS: q = u - alpha v_hat
//CGS: t = u + q
magma_zcgs_3(
b.num_rows,
b.num_cols,
C_alpha,
SpMV_out_1.dval+dofs,
C_u.dval,
C_q.dval,
SpMV_in_2.dval+dofs,
queue );
// BiCGSTAB: s = r - alpha v
magma_zbicgstab_2(
b.num_rows,
b.num_cols,
B_alpha,
B_r.dval,
SpMV_out_1.dval+2*dofs,
SpMV_in_2.dval+2*dofs,
queue );
// gather everything for a block-SpMV
//magma_zcopy( dofs, Q_q.dval, 1, SpMV_in_2.dval , 1, queue );
//magma_zcopy( dofs, C_t.dval, 1, SpMV_in_2.dval+dofs , 1, queue );
//magma_zcopy( dofs, B_s.dval, 1, SpMV_in_2.dval+2*dofs , 1, queue );
// block SpMV
CHECK( magma_z_spmv( c_one, A, SpMV_in_2, c_zero, SpMV_out_2, queue ));
// scatter results
//magma_zcopy( dofs, SpMV_out_2.dval , 1, Q_wt.dval, 1, queue );
//magma_zcopy( dofs, SpMV_out_2.dval+dofs , 1, C_rt.dval, 1, queue );
//magma_zcopy( dofs, SpMV_out_2.dval+2*dofs , 1, B_t.dval, 1, queue );
Q_rho1 = Q_rho;
if( mdot == 0 ) {
//QMR rho = norm(y);
Q_rho = magma_zsqrt( magma_zdotc( dofs, Q_y.dval, 1, Q_y.dval, 1), queue );
// BiCGSTAB
B_omega = magma_zdotc( dofs, SpMV_out_2.dval+2*dofs, 1, SpMV_in_2.dval+2*dofs, 1 ) // omega = <s,t>/<t,t>
/ magma_zdotc( dofs, SpMV_out_2.dval+2*dofs, 1, SpMV_out_2.dval+2*dofs, 1, queue );
}else{
magma_zmdotc4(
b.num_rows,
Q_y.dval,
Q_y.dval,
Q_y.dval,
Q_y.dval,
SpMV_out_2.dval+2*dofs,
SpMV_in_2.dval+2*dofs,
SpMV_out_2.dval+2*dofs,
SpMV_out_2.dval+2*dofs,
d1.dval,
d2.dval,
skp.val,
queue );
Q_rho = magma_zsqrt(skp.val[0]);
B_omega = skp.val[2]/skp.val[3];
}
// QMR
Q_thet1 = Q_thet;
Q_thet = Q_rho / (Q_gamm * MAGMA_Z_MAKE( MAGMA_Z_ABS(Q_beta), 0.0 ));
Q_gamm1 = Q_gamm;
Q_gamm = c_one / magma_zsqrt(c_one + Q_thet*Q_thet);
Q_eta = - Q_eta * Q_rho1 * Q_gamm * Q_gamm / (Q_beta * Q_gamm1 * Q_gamm1);
if( solver_par->numiter == 1 ){
//QMR: d = eta * p + pds * d;
//QMR: s = eta * pt + pds * d;
//QMR: x = x + d;
//QMR: r = r - s;
magma_zqmr_4(
b.num_rows,
b.num_cols,
Q_eta,
SpMV_in_1.dval,
SpMV_out_1.dval,
Q_d.dval,
Q_s.dval,
Q_x.dval,
Q_r.dval,
queue );
}
else{
Q_pds = (Q_thet1 * Q_gamm) * (Q_thet1 * Q_gamm);
//QMR: d = eta * p + pds * d;
//QMR: s = eta * pt + pds * d;
//QMR: x = x + d;
//QMR: r = r - s;
magma_zqmr_5(
b.num_rows,
b.num_cols,
Q_eta,
Q_pds,
SpMV_in_1.dval,
SpMV_out_1.dval,
Q_d.dval,
Q_s.dval,
Q_x.dval,
Q_r.dval,
queue );
}
// CGS: r = r -alpha*A u_hat
// CGS: x = x + alpha u_hat
magma_zcgs_4(
b.num_rows,
b.num_cols,
C_alpha,
SpMV_in_2.dval+dofs,
SpMV_out_2.dval+dofs,
C_x.dval,
C_r.dval,
queue );
// BiCGSTAB: x = x + alpha * p + omega * s
// BiCGSTAB: r = s - omega * t
magma_zbicgstab_3(
b.num_rows,
b.num_cols,
B_alpha,
B_omega,
SpMV_in_1.dval+2*dofs,
SpMV_in_2.dval+2*dofs,
SpMV_out_2.dval+2*dofs,
B_x.dval,
B_r.dval,
queue );
if( mdot == 0 ){
Q_res = magma_dznrm2( dofs, Q_r.dval, 1, queue );
C_res = magma_dznrm2( dofs, C_r.dval, 1, queue );
B_res = magma_dznrm2( dofs, B_r.dval, 1, queue );
//QMR: psi = norm(z);
Q_psi = magma_zsqrt( magma_zdotc( dofs, Q_z.dval, 1, Q_z.dval, 1), queue );
}else{
magma_zmdotc4(
b.num_rows,
Q_r.dval,
Q_r.dval,
C_r.dval,
C_r.dval,
B_r.dval,
B_r.dval,
Q_z.dval,
Q_z.dval,
d1.dval,
d2.dval,
skp.val,
queue );
Q_res = MAGMA_Z_ABS(magma_zsqrt(skp.val[0]));
C_res = MAGMA_Z_ABS(magma_zsqrt(skp.val[1]));
B_res = MAGMA_Z_ABS(magma_zsqrt(skp.val[2]));
//QMR: psi = norm(z);
Q_psi = magma_zsqrt(skp.val[3]);
}
//QMR: v = y / rho
//QMR: y = y / rho
//QMR: w = wt / psi
//QMR: z = z / psi
//QMR: wt = A' * q - beta' * w
//QMR: no precond: z = wt
magma_zqmr_6(
b.num_rows,
b.num_cols,
Q_beta,
Q_rho,
Q_psi,
Q_y.dval,
Q_z.dval,
Q_v.dval,
Q_w.dval,
SpMV_out_2.dval,
queue );
// printf(" %e %e %e\n", Q_res, C_res, B_res);
if( Q_res < res ){
res = Q_res;
flag = 1;
}
if( C_res < res ){
res = C_res;
flag = 2;
}
if( B_res < res ){
res = B_res;
flag = 3;
}
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;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
// copy back the best solver
switch ( flag ) {
case 1:
printf("%% QMR fastest solver.\n");
magma_zcopy( dofs, Q_x.dval, 1, x->dval, 1, queue );
break;
case 2:
printf("%% CGS fastest solver.\n");
magma_zcopy( dofs, C_x.dval, 1, x->dval, 1, queue );
break;
case 3:
printf("%% BiCGSTAB fastest solver.\n");
magma_zcopy( dofs, B_x.dval, 1, x->dval, 1, queue );
break;
}
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r_tld, &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 == 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_tld, queue );
magma_zmfree(&SpMV_in_1, queue );
magma_zmfree(&SpMV_out_1, queue );
magma_zmfree(&SpMV_in_2, queue );
magma_zmfree(&SpMV_out_2, queue );
magma_zmfree(&d1, queue );
magma_zmfree(&d2, queue );
magma_zmfree(&skp, queue );
magma_zmfree(&AT, queue );
// QMR
magma_zmfree(&Q_r, queue );
magma_zmfree(&Q_v, queue );
magma_zmfree(&Q_w, queue );
magma_zmfree(&Q_wt, queue );
magma_zmfree(&Q_d, queue );
magma_zmfree(&Q_s, queue );
magma_zmfree(&Q_z, queue );
magma_zmfree(&Q_q, queue );
magma_zmfree(&Q_p, queue );
magma_zmfree(&Q_pt, queue );
magma_zmfree(&Q_y, queue );
magma_zmfree(&Q_x, queue );
// CGS
magma_zmfree(&C_r, queue );
magma_zmfree(&C_rt, queue );
magma_zmfree(&C_x, queue );
magma_zmfree(&C_p, queue );
magma_zmfree(&C_q, queue );
magma_zmfree(&C_u, queue );
magma_zmfree(&C_v, queue );
magma_zmfree(&C_t, queue );
magma_zmfree(&C_p_hat, queue );
magma_zmfree(&C_q_hat, queue );
magma_zmfree(&C_u_hat, queue );
magma_zmfree(&C_v_hat, queue );
// BiCGSTAB
magma_zmfree(&B_r, queue );
magma_zmfree(&B_x, queue );
magma_zmfree(&B_p, queue );
magma_zmfree(&B_v, queue );
magma_zmfree(&B_s, queue );
magma_zmfree(&B_t, queue );
// destroy asynchronous queues
magma_queue_destroy( queues[0] );
if ( q1flag == 1 ) {
magma_queue_destroy( queues[1] );
}
magma_queue_destroy( queues[2] );
solver_par->info = info;
return info;
} /* magma_zbombard_merge */
