| 1 |
// @HEADER |
| 2 |
// ***************************************************************************** |
| 3 |
// Tpetra: Templated Linear Algebra Services Package |
| 4 |
// |
| 5 |
// Copyright 2008 NTESS and the Tpetra contributors. |
| 6 |
// SPDX-License-Identifier: BSD-3-Clause |
| 7 |
// ***************************************************************************** |
| 8 |
// @HEADER |
| 9 |
|
| 10 |
#ifndef TPETRA_TSQR_ADAPTOR_UQ_PCE_HPP |
| 11 |
#define TPETRA_TSQR_ADAPTOR_UQ_PCE_HPP |
| 12 |
|
| 13 |
#include <Tpetra_ConfigDefs.hpp> // HAVE_TPETRA_TSQR, etc. |
| 14 |
|
| 15 |
#ifdef HAVE_TPETRA_TSQR |
| 16 |
|
| 17 |
#include "Stokhos_Sacado_Kokkos_UQ_PCE.hpp" |
| 18 |
|
| 19 |
# include "Tsqr_NodeTsqrFactory.hpp" // create intranode TSQR object |
| 20 |
# include "Tsqr.hpp" // full (internode + intranode) TSQR |
| 21 |
# include "Tsqr_DistTsqr.hpp" // internode TSQR |
| 22 |
// Subclass of TSQR::MessengerBase, implemented using Teuchos |
| 23 |
// communicator template helper functions |
| 24 |
# include "Tsqr_TeuchosMessenger.hpp" |
| 25 |
# include "Tpetra_MultiVector.hpp" |
| 26 |
# include "Teuchos_ParameterListAcceptorDefaultBase.hpp" |
| 27 |
# include <stdexcept> |
| 28 |
|
| 29 |
// Base TsqrAdator template we will specialize |
| 30 |
# include "Tpetra_TsqrAdaptor.hpp" |
| 31 |
|
| 32 |
namespace Tpetra { |
| 33 |
|
| 34 |
/// \class TsqrAdaptor |
| 35 |
/// \brief Adaptor from Tpetra::MultiVector to TSQR for UQ::PCE scalar type |
| 36 |
/// \author Eric Phipps |
| 37 |
/// |
| 38 |
/// This specialization works be extracting the underlying array within the |
| 39 |
/// multivector and converting to a standard scalar type. |
| 40 |
template <class Storage, class LO, class GO, class Node> |
| 41 |
class TsqrAdaptor< Tpetra::MultiVector< Sacado::UQ::PCE<Storage>, |
| 42 |
LO, GO, Node > > : |
| 43 |
public Teuchos::ParameterListAcceptorDefaultBase { |
| 44 |
public: |
| 45 |
typedef Tpetra::MultiVector< Sacado::UQ::PCE<Storage>, LO, GO, Node > MV; |
| 46 |
typedef typename MV::scalar_type mp_scalar_type; |
| 47 |
|
| 48 |
// For Sacado::UQ::PCE< Storage<Ordinal,Scalar,Device> > this is Scalar |
| 49 |
typedef typename mp_scalar_type::scalar_type scalar_type; |
| 50 |
typedef typename mp_scalar_type::ordinal_type mp_ordinal_type; |
| 51 |
typedef typename MV::local_ordinal_type ordinal_type; |
| 52 |
typedef Teuchos::SerialDenseMatrix<ordinal_type, scalar_type> dense_matrix_type; |
| 53 |
typedef typename Teuchos::ScalarTraits<scalar_type>::magnitudeType magnitude_type; |
| 54 |
|
| 55 |
private: |
| 56 |
using node_tsqr_factory_type = |
| 57 |
TSQR::NodeTsqrFactory<scalar_type, ordinal_type, |
| 58 |
typename MV::device_type>; |
| 59 |
using node_tsqr_type = TSQR::NodeTsqr<ordinal_type, scalar_type>; |
| 60 |
using dist_tsqr_type = TSQR::DistTsqr<ordinal_type, scalar_type>; |
| 61 |
using tsqr_type = TSQR::Tsqr<ordinal_type, scalar_type>; |
| 62 |
|
| 63 |
public: |
| 64 |
/// \brief Constructor (that accepts a parameter list). |
| 65 |
/// |
| 66 |
/// \param plist [in/out] List of parameters for configuring TSQR. |
| 67 |
/// The specific parameter keys that are read depend on the TSQR |
| 68 |
/// implementation. For details, call \c getValidParameters() |
| 69 |
/// and examine the documentation embedded therein. |
| 70 |
TsqrAdaptor (const Teuchos::RCP<Teuchos::ParameterList>& plist) : |
| 71 |
nodeTsqr_ (node_tsqr_factory_type::getNodeTsqr ()), |
| 72 |
distTsqr_ (new dist_tsqr_type), |
| 73 |
tsqr_ (new tsqr_type (nodeTsqr_, distTsqr_)), |
| 74 |
ready_ (false) |
| 75 |
{ |
| 76 |
setParameterList (plist); |
| 77 |
} |
| 78 |
|
| 79 |
//! Constructor (that uses default parameters). |
| 80 |
TsqrAdaptor () : |
| 81 |
nodeTsqr_ (new node_tsqr_factory_type::getNodeTsqr ()), |
| 82 |
distTsqr_ (new dist_tsqr_type), |
| 83 |
tsqr_ (new tsqr_type (nodeTsqr_, distTsqr_)), |
| 84 |
ready_ (false) |
| 85 |
{ |
| 86 |
setParameterList (Teuchos::null); |
| 87 |
} |
| 88 |
|
| 89 |
Teuchos::RCP<const Teuchos::ParameterList> |
| 90 |
getValidParameters () const |
| 91 |
{ |
| 92 |
using Teuchos::RCP; |
| 93 |
using Teuchos::rcp; |
| 94 |
using Teuchos::ParameterList; |
| 95 |
using Teuchos::parameterList; |
| 96 |
|
| 97 |
if (defaultParams_.is_null()) { |
| 98 |
RCP<ParameterList> params = parameterList ("TSQR implementation"); |
| 99 |
params->set ("NodeTsqr", *(nodeTsqr_->getValidParameters ())); |
| 100 |
params->set ("DistTsqr", *(distTsqr_->getValidParameters ())); |
| 101 |
defaultParams_ = params; |
| 102 |
} |
| 103 |
return defaultParams_; |
| 104 |
} |
| 105 |
|
| 106 |
void |
| 107 |
setParameterList (const Teuchos::RCP<Teuchos::ParameterList>& plist) |
| 108 |
{ |
| 109 |
using Teuchos::ParameterList; |
| 110 |
using Teuchos::parameterList; |
| 111 |
using Teuchos::RCP; |
| 112 |
using Teuchos::sublist; |
| 113 |
|
| 114 |
RCP<ParameterList> params = plist.is_null() ? |
| 115 |
parameterList (*getValidParameters ()) : plist; |
| 116 |
nodeTsqr_->setParameterList (sublist (params, "NodeTsqr")); |
| 117 |
distTsqr_->setParameterList (sublist (params, "DistTsqr")); |
| 118 |
|
| 119 |
this->setMyParamList (params); |
| 120 |
} |
| 121 |
|
| 122 |
/// \brief Compute QR factorization [Q,R] = qr(A,0). |
| 123 |
/// |
| 124 |
/// \param A [in/out] On input: the multivector to factor. |
| 125 |
/// Overwritten with garbage on output. |
| 126 |
/// |
| 127 |
/// \param Q [out] On output: the (explicitly stored) Q factor in |
| 128 |
/// the QR factorization of the (input) multivector A. |
| 129 |
/// |
| 130 |
/// \param R [out] On output: the R factor in the QR factorization |
| 131 |
/// of the (input) multivector A. |
| 132 |
/// |
| 133 |
/// \param forceNonnegativeDiagonal [in] If true, then (if |
| 134 |
/// necessary) do extra work (modifying both the Q and R |
| 135 |
/// factors) in order to force the R factor to have a |
| 136 |
/// nonnegative diagonal. |
| 137 |
/// |
| 138 |
/// \warning Currently, this method only works if A and Q have the |
| 139 |
/// same communicator and row distribution ("Map," in Petra |
| 140 |
/// terms) as those of the multivector given to this adapter |
| 141 |
/// instance's constructor. Otherwise, the result of this |
| 142 |
/// method is undefined. |
| 143 |
void |
| 144 |
factorExplicit (MV& A, |
| 145 |
MV& Q, |
| 146 |
dense_matrix_type& R, |
| 147 |
const bool forceNonnegativeDiagonal=false) |
| 148 |
{ |
| 149 |
prepareTsqr (Q); // Finish initializing TSQR. |
| 150 |
|
| 151 |
ordinal_type numRows; |
| 152 |
ordinal_type numCols; |
| 153 |
ordinal_type LDA; |
| 154 |
ordinal_type LDQ; |
| 155 |
scalar_type* A_ptr; |
| 156 |
scalar_type* Q_ptr; |
| 157 |
|
| 158 |
getNonConstView (numRows, numCols, A_ptr, LDA, A); |
| 159 |
getNonConstView (numRows, numCols, Q_ptr, LDQ, Q); |
| 160 |
const bool contiguousCacheBlocks = false; |
| 161 |
tsqr_->factorExplicitRaw (numRows, numCols, A_ptr, LDA, |
| 162 |
Q_ptr, LDQ, R.values (), R.stride (), |
| 163 |
contiguousCacheBlocks, |
| 164 |
forceNonnegativeDiagonal); |
| 165 |
} |
| 166 |
|
| 167 |
/// \brief Rank-revealing decomposition |
| 168 |
/// |
| 169 |
/// Using the R factor and explicit Q factor from |
| 170 |
/// factorExplicit(), compute the singular value decomposition |
| 171 |
/// (SVD) of R: \f$R = U \Sigma V^*\f$. If R is full rank (with |
| 172 |
/// respect to the given relative tolerance \c tol), do not modify |
| 173 |
/// Q or R. Otherwise, compute \f$Q := Q \cdot U\f$ and \f$R := |
| 174 |
/// \Sigma V^*\f$ in place. If R was modified, then it may not |
| 175 |
/// necessarily be upper triangular on output. |
| 176 |
/// |
| 177 |
/// \param Q [in/out] On input: explicit Q factor computed by |
| 178 |
/// factorExplicit(). (Must be an orthogonal resp. unitary |
| 179 |
/// matrix.) On output: If R is of full numerical rank with |
| 180 |
/// respect to the tolerance tol, Q is unmodified. Otherwise, Q |
| 181 |
/// is updated so that the first \c rank columns of Q are a |
| 182 |
/// basis for the column space of A (the original matrix whose |
| 183 |
/// QR factorization was computed by factorExplicit()). The |
| 184 |
/// remaining columns of Q are a basis for the null space of A. |
| 185 |
/// |
| 186 |
/// \param R [in/out] On input: N by N upper triangular matrix |
| 187 |
/// with leading dimension LDR >= N. On output: if input is |
| 188 |
/// full rank, R is unchanged on output. Otherwise, if \f$R = U |
| 189 |
/// \Sigma V^*\f$ is the SVD of R, on output R is overwritten |
| 190 |
/// with \f$\Sigma \cdot V^*\f$. This is also an N by N matrix, |
| 191 |
/// but it may not necessarily be upper triangular. |
| 192 |
/// |
| 193 |
/// \param tol [in] Relative tolerance for computing the numerical |
| 194 |
/// rank of the matrix R. |
| 195 |
/// |
| 196 |
/// \return Rank \f$r\f$ of R: \f$ 0 \leq r \leq N\f$. |
| 197 |
int |
| 198 |
revealRank (MV& Q, |
| 199 |
dense_matrix_type& R, |
| 200 |
const magnitude_type& tol) |
| 201 |
{ |
| 202 |
prepareTsqr (Q); // Finish initializing TSQR. |
| 203 |
|
| 204 |
// FIXME (mfh 18 Oct 2010) Check Teuchos::Comm<int> object in Q |
| 205 |
// to make sure it is the same communicator as the one we are |
| 206 |
// using in our dist_tsqr_type implementation. |
| 207 |
|
| 208 |
ordinal_type numRows; |
| 209 |
ordinal_type numCols; |
| 210 |
scalar_type* Q_ptr; |
| 211 |
ordinal_type LDQ; |
| 212 |
getNonConstView (numRows, numCols, Q_ptr, LDQ, Q); |
| 213 |
const bool contiguousCacheBlocks = false; |
| 214 |
return tsqr_->revealRankRaw (numRows, numCols, Q_ptr, LDQ, |
| 215 |
R.values (), R.stride (), tol, |
| 216 |
contiguousCacheBlocks); |
| 217 |
} |
| 218 |
|
| 219 |
private: |
| 220 |
//! The intranode TSQR implementation instance. |
| 221 |
Teuchos::RCP<node_tsqr_type> nodeTsqr_; |
| 222 |
|
| 223 |
//! The internode TSQR implementation instance. |
| 224 |
Teuchos::RCP<dist_tsqr_type> distTsqr_; |
| 225 |
|
| 226 |
//! The (full) TSQR implementation instance. |
| 227 |
Teuchos::RCP<tsqr_type> tsqr_; |
| 228 |
|
| 229 |
//! Default parameter list. Initialized by getValidParameters(). |
| 230 |
mutable Teuchos::RCP<const Teuchos::ParameterList> defaultParams_; |
| 231 |
|
| 232 |
//! Whether TSQR has been fully initialized. |
| 233 |
bool ready_; |
| 234 |
|
| 235 |
/// \brief Finish TSQR initialization. |
| 236 |
/// |
| 237 |
/// The intranode and internode TSQR implementations both have a |
| 238 |
/// two-stage initialization procedure: first, setting parameters |
| 239 |
/// (which may happen at construction), and second, getting |
| 240 |
/// information they need from the multivector input in order to |
| 241 |
/// finish initialization. For intranode TSQR, this includes the |
| 242 |
/// Kokkos Node instance; for internode TSQR, this includes the |
| 243 |
/// communicator. The second stage of initialization happens in |
| 244 |
/// this class' computational routines; all of those routines |
| 245 |
/// accept one or more multivector inputs, which this method can |
| 246 |
/// use for finishing initialization. Thus, users of this class |
| 247 |
/// never need to see the two-stage initialization. |
| 248 |
/// |
| 249 |
/// \param mv [in] Multivector object, used only to access the |
| 250 |
/// underlying communicator object (in this case, |
| 251 |
/// Teuchos::Comm<int>, accessed via the Tpetra::Map belonging |
| 252 |
/// to the multivector) and Kokkos Node instance. All |
| 253 |
/// multivector objects used with this Adaptor instance must |
| 254 |
/// have the same map, communicator, and Kokkos Node instance. |
| 255 |
void |
| 256 |
prepareTsqr (const MV& mv) |
| 257 |
{ |
| 258 |
if (! ready_) { |
| 259 |
prepareDistTsqr (mv); |
| 260 |
ready_ = true; |
| 261 |
} |
| 262 |
} |
| 263 |
|
| 264 |
/// \brief Finish interprocess TSQR initialization. |
| 265 |
/// |
| 266 |
/// \param mv [in] A valid Tpetra::MultiVector instance whose |
| 267 |
/// communicator wrapper we will use to prepare TSQR. |
| 268 |
/// |
| 269 |
/// \note It's OK to call this method more than once; it is idempotent. |
| 270 |
void |
| 271 |
prepareDistTsqr (const MV& mv) |
| 272 |
{ |
| 273 |
using Teuchos::RCP; |
| 274 |
using Teuchos::rcp_implicit_cast; |
| 275 |
typedef TSQR::TeuchosMessenger<scalar_type> mess_type; |
| 276 |
typedef TSQR::MessengerBase<scalar_type> base_mess_type; |
| 277 |
|
| 278 |
RCP<const Teuchos::Comm<int> > comm = mv.getMap()->getComm(); |
| 279 |
RCP<mess_type> mess (new mess_type (comm)); |
| 280 |
RCP<base_mess_type> messBase = rcp_implicit_cast<base_mess_type> (mess); |
| 281 |
distTsqr_->init (messBase); |
| 282 |
} |
| 283 |
|
| 284 |
/// \brief Extract a nonpersistent view of A's data as a |
| 285 |
/// scalar_type matrix, stored as a flat column-major array. |
| 286 |
/// |
| 287 |
/// \warning TSQR does not currently support multivectors with |
| 288 |
/// nonconstant stride. If A has nonconstant stride, this |
| 289 |
/// method will throw an exception. |
| 290 |
static void |
| 291 |
getNonConstView (ordinal_type& numRows, |
| 292 |
ordinal_type& numCols, |
| 293 |
scalar_type*& A_ptr, |
| 294 |
ordinal_type& LDA, |
| 295 |
const MV& A) |
| 296 |
{ |
| 297 |
// FIXME (mfh 25 Oct 2010) We should be able to run TSQR even if |
| 298 |
// storage of A uses nonconstant stride internally. We would |
| 299 |
// have to copy and pack into a matrix with constant stride, and |
| 300 |
// then unpack on exit. For now we choose just to raise an |
| 301 |
// exception. |
| 302 |
TEUCHOS_TEST_FOR_EXCEPTION |
| 303 |
(! A.isConstantStride(), std::invalid_argument, |
| 304 |
"TSQR does not currently support Tpetra::MultiVector " |
| 305 |
"inputs that do not have constant stride."); |
| 306 |
|
| 307 |
// FIXME (mfh 16 Jan 2016) When I got here, I found strides[0] |
| 308 |
// instead of strides[1] for the stride. I don't think this is |
| 309 |
// right. However, I don't know about these Stokhos scalar |
| 310 |
// types so I'll just do what was here. |
| 311 |
// |
| 312 |
// STOKHOS' TYPES ARE NOT TESTED WITH TSQR REGULARLY SO IT IS |
| 313 |
// POSSIBLE THAT THE ORIGINAL CODE WAS WRONG. |
| 314 |
|
| 315 |
typedef typename MV::dual_view_type view_type; |
| 316 |
typedef typename view_type::t_dev::array_type flat_array_type; |
| 317 |
|
| 318 |
// Reinterpret the data as a longer array of the base scalar |
| 319 |
// type. TSQR currently forbids MultiVector input with |
| 320 |
// nonconstant stride, so we need not worry about that here. |
| 321 |
|
| 322 |
flat_array_type flat_mv = A.getLocalViewDevice(); |
| 323 |
|
| 324 |
numRows = static_cast<ordinal_type> (flat_mv.extent(0)); |
| 325 |
numCols = static_cast<ordinal_type> (flat_mv.extent(1)); |
| 326 |
A_ptr = flat_mv.data (); |
| 327 |
|
| 328 |
ordinal_type strides[2]; |
| 329 |
flat_mv.stride (strides); |
| 330 |
LDA = strides[0]; |
| 331 |
} |
| 332 |
}; |
| 333 |
} // namespace Tpetra |
| 334 |
|
| 335 |
#endif // HAVE_TPETRA_TSQR |
| 336 |
|
| 337 |
#endif // TPETRA_TSQR_ADAPTOR_UQ_PCE_HPP |