Amesos2 - Direct Sparse Solver Interfaces  Version of the Day
Amesos2_Tacho_def.hpp
1 // @HEADER
2 //
3 // ***********************************************************************
4 //
5 // Amesos2: Templated Direct Sparse Solver Package
6 // Copyright 2011 Sandia Corporation
7 //
8 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
9 // the U.S. Government retains certain rights in this software.
10 //
11 // Redistribution and use in source and binary forms, with or without
12 // modification, are permitted provided that the following conditions are
13 // met:
14 //
15 // 1. Redistributions of source code must retain the above copyright
16 // notice, this list of conditions and the following disclaimer.
17 //
18 // 2. Redistributions in binary form must reproduce the above copyright
19 // notice, this list of conditions and the following disclaimer in the
20 // documentation and/or other materials provided with the distribution.
21 //
22 // 3. Neither the name of the Corporation nor the names of the
23 // contributors may be used to endorse or promote products derived from
24 // this software without specific prior written permission.
25 //
26 // THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
27 // EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29 // PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
30 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
31 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
32 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
33 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
34 // LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
35 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
36 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37 //
38 // Questions? Contact Sivasankaran Rajamanickam (srajama@sandia.gov)
39 //
40 // ***********************************************************************
41 //
42 // @HEADER
43 
44 #ifndef AMESOS2_TACHO_DEF_HPP
45 #define AMESOS2_TACHO_DEF_HPP
46 
47 #include <Teuchos_Tuple.hpp>
48 #include <Teuchos_ParameterList.hpp>
49 #include <Teuchos_StandardParameterEntryValidators.hpp>
50 
52 #include "Amesos2_Tacho_decl.hpp"
53 #include "Amesos2_Util.hpp"
54 
55 namespace Amesos2 {
56 
57 template <class Matrix, class Vector>
59  Teuchos::RCP<const Matrix> A,
60  Teuchos::RCP<Vector> X,
61  Teuchos::RCP<const Vector> B )
62  : SolverCore<Amesos2::TachoSolver,Matrix,Vector>(A, X, B)
63 {
64 }
65 
66 
67 template <class Matrix, class Vector>
69 {
70 }
71 
72 template <class Matrix, class Vector>
73 std::string
75 {
76  std::ostringstream oss;
77  oss << "Tacho solver interface";
78  return oss.str();
79 }
80 
81 template<class Matrix, class Vector>
82 int
84 {
85  return(0);
86 }
87 
88 template <class Matrix, class Vector>
89 int
91 {
92  int status = 0;
93  if ( this->root_ ) {
94  if(do_optimization()) {
95  this->matrixA_->returnRowPtr_kokkos_view(host_row_ptr_view_);
96  this->matrixA_->returnColInd_kokkos_view(host_cols_view_);
97  }
98 
99  // TODO: Confirm param options
100  // data_.solver.setMaxNumberOfSuperblocks(data_.max_num_superblocks);
101 
102  // Symbolic factorization currently must be done on host
103  data_.solver.analyze(this->globalNumCols_, host_row_ptr_view_, host_cols_view_);
104  }
105  return status;
106 }
107 
108 
109 template <class Matrix, class Vector>
110 int
112 {
113  int status = 0;
114  if ( this->root_ ) {
115  if(do_optimization()) {
116  this->matrixA_->returnValues_kokkos_view(device_nzvals_view_);
117  }
118  data_.solver.factorize(device_nzvals_view_);
119  }
120  return status;
121 }
122 
123 template <class Matrix, class Vector>
124 int
126  const Teuchos::Ptr<const MultiVecAdapter<Vector> > B) const
127 {
128  using Teuchos::as;
129 
130  const global_size_type ld_rhs = this->root_ ? X->getGlobalLength() : 0;
131  const size_t nrhs = X->getGlobalNumVectors();
132 
133  // don't allocate b since it's handled by the copy manager and might just be
134  // be assigned, not copied anyways.
135  // also don't allocate x since we will also use do_get to allocate this if
136  // necessary. When a copy is not necessary we'll solve directly to the x
137  // values in the MV.
138  { // Get values from RHS B
139 #ifdef HAVE_AMESOS2_TIMERS
140  Teuchos::TimeMonitor mvConvTimer(this->timers_.vecConvTime_);
141  Teuchos::TimeMonitor redistTimer(this->timers_.vecRedistTime_);
142 #endif
143  Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
144  device_solve_array_t>::do_get(B, this->bValues_,
145  as<size_t>(ld_rhs),
146  ROOTED, this->rowIndexBase_);
147 
148  // If it's not a match and we copy instead of ptr assignement, we will
149  // copy the x values here when we just wanted to get uninitialized space.
150  // MDM-DISCUSS Need to decide how to request the underlying API to learn
151  // whether we will need to copy or not.
152  Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
153  device_solve_array_t>::do_get(X, this->xValues_,
154  as<size_t>(ld_rhs),
155  ROOTED, this->rowIndexBase_);
156  }
157 
158  int ierr = 0; // returned error code
159 
160  if ( this->root_ ) { // Do solve!
161 #ifdef HAVE_AMESOS2_TIMER
162  Teuchos::TimeMonitor solveTimer(this->timers_.solveTime_);
163 #endif
164  // Bump up the workspace size if needed
165  if (workspace_.extent(0) < this->globalNumRows_ || workspace_.extent(1) < nrhs) {
166  workspace_ = device_solve_array_t(
167  Kokkos::ViewAllocateWithoutInitializing("t"), this->globalNumRows_, nrhs);
168  }
169 
170  data_.solver.solve(xValues_, bValues_, workspace_);
171 
172  int status = 0; // TODO: determine what error handling will be
173  if(status != 0) {
174  ierr = status;
175  }
176  }
177 
178  /* All processes should have the same error code */
179  Teuchos::broadcast(*(this->getComm()), 0, &ierr);
180 
181  TEUCHOS_TEST_FOR_EXCEPTION( ierr != 0, std::runtime_error,
182  "tacho_solve has error code: " << ierr );
183 
184  /* Update X's global values */
185  {
186 #ifdef HAVE_AMESOS2_TIMERS
187  Teuchos::TimeMonitor redistTimer(this->timers_.vecRedistTime_);
188 #endif
189 
190  // This will do nothing is if the target view matches the src view, which
191  // can be the case if the memory spaces match. See comments above for do_get.
192  Util::template put_1d_data_helper_kokkos_view<
193  MultiVecAdapter<Vector>,device_solve_array_t>::do_put(X, xValues_,
194  as<size_t>(ld_rhs),
195  ROOTED, this->rowIndexBase_);
196  }
197 
198  return(ierr);
199 }
200 
201 
202 template <class Matrix, class Vector>
203 bool
205 {
206  // Tacho can only apply the solve routines to square matrices
207  return( this->matrixA_->getGlobalNumRows() == this->matrixA_->getGlobalNumCols() );
208 }
209 
210 
211 template <class Matrix, class Vector>
212 void
213 TachoSolver<Matrix,Vector>::setParameters_impl(const Teuchos::RCP<Teuchos::ParameterList> & parameterList )
214 {
215  RCP<const Teuchos::ParameterList> valid_params = getValidParameters_impl();
216 
217  // TODO: Confirm param options
218  // data_.num_kokkos_threads = parameterList->get<int>("kokkos-threads", 1);
219  // data_.max_num_superblocks = parameterList->get<int>("max-num-superblocks", 4);
220 }
221 
222 
223 template <class Matrix, class Vector>
224 Teuchos::RCP<const Teuchos::ParameterList>
226 {
227  static Teuchos::RCP<const Teuchos::ParameterList> valid_params;
228 
229  if( is_null(valid_params) ){
230  Teuchos::RCP<Teuchos::ParameterList> pl = Teuchos::parameterList();
231 
232  // TODO: Confirm param options
233  // pl->set("kokkos-threads", 1, "Number of threads");
234  // pl->set("max-num-superblocks", 4, "Max number of superblocks");
235 
236  valid_params = pl;
237  }
238 
239  return valid_params;
240 }
241 
242 template <class Matrix, class Vector>
243 bool
245  return (this->root_ && (this->matrixA_->getComm()->getSize() == 1));
246 }
247 
248 template <class Matrix, class Vector>
249 bool
251 {
252 
253  if(current_phase == SOLVE) {
254  return(false);
255  }
256 
257  if(!do_optimization()) {
258 #ifdef HAVE_AMESOS2_TIMERS
259  Teuchos::TimeMonitor convTimer(this->timers_.mtxConvTime_);
260 #endif
261 
262  // Note views are allocated but eventually we should remove this.
263  // The internal copy manager will decide if we can assign or deep_copy
264  // and then allocate if necessary. However the GPU solvers are serial right
265  // now so I didn't complete refactoring the matrix code for the parallel
266  // case. If we added that later, we should have it hooked up to the copy
267  // manager and then these allocations can go away.
268  if( this->root_ ) {
269  device_nzvals_view_ = device_value_type_array(
270  Kokkos::ViewAllocateWithoutInitializing("nzvals"), this->globalNumNonZeros_);
271  host_cols_view_ = host_ordinal_type_array(
272  Kokkos::ViewAllocateWithoutInitializing("colind"), this->globalNumNonZeros_);
273  host_row_ptr_view_ = host_size_type_array(
274  Kokkos::ViewAllocateWithoutInitializing("rowptr"), this->globalNumRows_ + 1);
275  }
276 
277  typename host_size_type_array::value_type nnz_ret = 0;
278  {
279  #ifdef HAVE_AMESOS2_TIMERS
280  Teuchos::TimeMonitor mtxRedistTimer( this->timers_.mtxRedistTime_ );
281  #endif
282 
283  TEUCHOS_TEST_FOR_EXCEPTION( this->rowIndexBase_ != this->columnIndexBase_,
284  std::runtime_error,
285  "Row and column maps have different indexbase ");
286 
287  Util::get_crs_helper_kokkos_view<MatrixAdapter<Matrix>,
288  device_value_type_array, host_ordinal_type_array, host_size_type_array>::do_get(
289  this->matrixA_.ptr(),
290  device_nzvals_view_,
291  host_cols_view_,
292  host_row_ptr_view_,
293  nnz_ret,
294  ROOTED, ARBITRARY,
295  this->columnIndexBase_);
296  }
297  }
298 
299  return true;
300 }
301 
302 
303 template<class Matrix, class Vector>
304 const char* TachoSolver<Matrix,Vector>::name = "Tacho";
305 
306 
307 } // end namespace Amesos2
308 
309 #endif // AMESOS2_TACHO_DEF_HPP
Amesos2::SolverCore: A templated interface for interaction with third-party direct sparse solvers...
Definition: Amesos2_SolverCore_decl.hpp:105
EPhase
Used to indicate a phase in the direct solution.
Definition: Amesos2_TypeDecl.hpp:65
Amesos2 interface to the Tacho package.
Definition: Amesos2_Tacho_decl.hpp:67
int symbolicFactorization_impl()
Perform symbolic factorization of the matrix using Tacho.
Definition: Amesos2_Tacho_def.hpp:90
std::string description() const
Returns a short description of this Solver.
Definition: Amesos2_Tacho_def.hpp:74
Utility functions for Amesos2.
bool matrixShapeOK_impl() const
Determines whether the shape of the matrix is OK for this solver.
Definition: Amesos2_Tacho_def.hpp:204
int numericFactorization_impl()
Tacho specific numeric factorization.
Definition: Amesos2_Tacho_def.hpp:111
int preOrdering_impl()
Performs pre-ordering on the matrix to increase efficiency.
Definition: Amesos2_Tacho_def.hpp:83
Teuchos::RCP< const Teuchos::ParameterList > getValidParameters_impl() const
Definition: Amesos2_Tacho_def.hpp:225
int solve_impl(const Teuchos::Ptr< MultiVecAdapter< Vector > > X, const Teuchos::Ptr< const MultiVecAdapter< Vector > > B) const
Tacho specific solve.
Definition: Amesos2_Tacho_def.hpp:125
TachoSolver(Teuchos::RCP< const Matrix > A, Teuchos::RCP< Vector > X, Teuchos::RCP< const Vector > B)
Initialize from Teuchos::RCP.
Definition: Amesos2_Tacho_def.hpp:58
bool loadA_impl(EPhase current_phase)
Reads matrix data into internal structures.
Definition: Amesos2_Tacho_def.hpp:250
bool do_optimization() const
can we optimize size_type and ordinal_type for straight pass through
Definition: Amesos2_Tacho_def.hpp:244
A templated MultiVector class adapter for Amesos2.
Definition: Amesos2_MultiVecAdapter_decl.hpp:176
~TachoSolver()
Destructor.
Definition: Amesos2_Tacho_def.hpp:68