libtrilinos-tpetra-dev 12.10.1-3 / usr / include / trilinos / Tsqr_TsqrTest.hpp

//@HEADER
// ************************************************************************
//
//          Kokkos: Node API and Parallel Node Kernels
//              Copyright (2008) Sandia Corporation
//
// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
// the U.S. Government retains certain rights in this software.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Questions? Contact Michael A. Heroux (maherou@sandia.gov)
//
// ************************************************************************
//@HEADER

#ifndef __TSQR_Test_TsqrTest_hpp
#define __TSQR_Test_TsqrTest_hpp

#include <Tsqr.hpp>
#ifdef HAVE_KOKKOSTSQR_TBB
#  include <TbbTsqr.hpp>
#endif // HAVE_KOKKOSTSQR_TBB
#include <Tsqr_TestSetup.hpp>
#include <Tsqr_GlobalVerify.hpp>
#include <Tsqr_printGlobalMatrix.hpp>
#include <Tsqr_verifyTimerConcept.hpp>
#include <Teuchos_ScalarTraits.hpp>

#include <cstring> // size_t
#include <iostream>
#include <stdexcept>
#include <string>


namespace TSQR {
  namespace Test {

    template<class TsqrType>
    class TsqrVerifier {
    public:
      typedef TsqrType tsqr_type;
      typedef typename tsqr_type::scalar_type scalar_type;
      typedef typename tsqr_type::ordinal_type ordinal_type;
      typedef Matrix<ordinal_type, scalar_type> matrix_type;
      typedef typename tsqr_type::FactorOutput factor_output_type;
      typedef MessengerBase<scalar_type> messenger_type;
      typedef Teuchos::RCP<messenger_type> messenger_ptr;

      static void
      verify (tsqr_type& tsqr,
              const messenger_ptr& scalarComm,
              const matrix_type& A_local,
              matrix_type& A_copy,
              matrix_type& Q_local,
              matrix_type& R,
              const bool contiguousCacheBlocks,
              const bool b_debug = false)
      {
        using std::cerr;
        using std::endl;

        const ordinal_type nrows_local = A_local.nrows();
        const ordinal_type ncols = A_local.ncols();

        // If specified, rearrange cache blocks in the copy.
        if (contiguousCacheBlocks) {
          tsqr.cache_block (nrows_local, ncols, A_copy.get(),
                            A_local.get(), A_local.lda());
          if (b_debug) {
            scalarComm->barrier ();
            if (scalarComm->rank () == 0)
              cerr << "-- Cache-blocked input matrix to factor." << endl;
          }
        }
        else {
          deep_copy (A_copy, A_local);
        }

        const bool testFactorExplicit = true;
        if (testFactorExplicit) {
          tsqr.factorExplicit (A_copy.view(), Q_local.view(), R.view(),
                               contiguousCacheBlocks);
          if (b_debug) {
            scalarComm->barrier ();
            if (scalarComm->rank () == 0)
              cerr << "-- Finished Tsqr::factorExplicit" << endl;
          }
        }
        else {
          // Factor the (copy of the) matrix.
          factor_output_type factorOutput =
            tsqr.factor (nrows_local, ncols, A_copy.get(), A_copy.lda(),
                         R.get(), R.lda(), contiguousCacheBlocks);
          if (b_debug) {
            scalarComm->barrier ();
            if (scalarComm->rank () == 0)
              cerr << "-- Finished Tsqr::factor" << endl;
          }

          // Compute the explicit Q factor in Q_local
          tsqr.explicit_Q (nrows_local,
                           ncols, A_copy.get(), A_copy.lda(), factorOutput,
                           ncols, Q_local.get(), Q_local.lda(),
                           contiguousCacheBlocks);
          if (b_debug) {
            scalarComm->barrier ();
            if (scalarComm->rank () == 0)
              cerr << "-- Finished Tsqr::explicit_Q" << endl;
          }
        }

        // "Un"-cache-block the output, if contiguous cache blocks were
        // used.  This is only necessary because global_verify() doesn't
        // currently support contiguous cache blocks.
        if (contiguousCacheBlocks) {
          // We can use A_copy as scratch space for un-cache-blocking
          // Q_local, since we're done using A_copy for other things.
          tsqr.un_cache_block (nrows_local, ncols, A_copy.get(),
                               A_copy.lda(), Q_local.get());
          // Overwrite Q_local with the un-cache-blocked Q factor.
          deep_copy (Q_local, A_copy);

          if (b_debug) {
            scalarComm->barrier ();
            if (scalarComm->rank () == 0)
              cerr << "-- Un-cache-blocked output Q factor" << endl;
          }
        }
      }
    };

    /// \function verifyTsqr
    /// \brief Test and print to stdout the accuracy of parallel TSQR
    ///
    /// \param which [in] Valid values: "MpiTbbTSQR" (for TBB-parallel
    ///   node-level TSQR underneath MPI-parallel TSQR), "MpiSeqTSQR"
    ///   (for cache-blocked sequential node-level TSQR underneath
    ///   MPI-parallel TSQR)
    ///
    /// \param scalarTypeName [in] Name of the Scalar type
    ///
    /// \param generator [in/out] Normal(0,1) (pseudo)random number
    ///   generator.  Only touched on MPI process 0.  Used to generate
    ///   random test matrices for the factorization.
    ///
    /// \param nrows_global [in] Number of rows in the entire test
    ///   matrix (over all processes) to generate.  The matrix will be
    ///   divided up in blocks of contiguous rows among the processes.
    ///
    /// \param ncols [in] Number of columns in the test matrix to
    ///   generate.
    ///
    /// \param ordinalComm [in/out] Object for communicating Ordinal
    ///   (integer index) objects among the processes
    ///
    /// \param scalarComm [in/out] Object for communicating Scalar
    ///   (matrix data) objects among the processes
    ///
    /// \param num_cores [in] Number of cores to use per MPI process
    ///   for Intel TBB parallelism within that process
    ///
    /// \param cache_size_hint [in] Cache size hint (per core) in
    ///   bytes.  If zero, a sensible default is used.
    ///
    /// \param contiguousCacheBlocks [in] Whether cache blocks
    ///   should be stored contiguously
    ///
    /// \param printFieldNames [in] Whether to print field names (only
    ///   appliable if not human_readable)
    ///
    /// \param human_readable [in] Whether output should be human
    ///   readable, or machine parseable
    ///
    /// \param b_debug [in] Whether to print debug output
    ///
    template<class Ordinal, class Scalar, class Generator>
    void
    verifyTsqr (const std::string& which,
                const std::string& scalarTypeName,
                Generator& generator,
                const Ordinal nrows_global,
                const Ordinal ncols,
                const Teuchos::RCP< MessengerBase< Ordinal > >& ordinalComm,
                const Teuchos::RCP< MessengerBase< Scalar > >& scalarComm,
                const int num_cores = 1,
                const size_t cache_size_hint = 0,
                const bool contiguousCacheBlocks,
                const bool printFieldNames,
                const bool human_readable = false,
                const bool b_debug = false)
    {
      typedef typename Teuchos::ScalarTraits<Scalar>::magnitudeType magnitude_type;
      using std::cerr;
      using std::cout;
      using std::endl;

      const bool b_extra_debug = false;
      const int nprocs = scalarComm->size();
      const int my_rank = scalarComm->rank();
      if (b_debug) {
        scalarComm->barrier ();
        if (my_rank == 0) {
          cerr << "tsqr_verify:" << endl;
        }
        scalarComm->barrier ();
      }
      const Ordinal nrows_local = numLocalRows (nrows_global, my_rank, nprocs);

      // Set up storage for the test problem.
      Matrix< Ordinal, Scalar > A_local (nrows_local, ncols);
      Matrix< Ordinal, Scalar > Q_local (nrows_local, ncols);
      if (std::numeric_limits<Scalar>::has_quiet_NaN) {
          A_local.fill (std::numeric_limits<Scalar>::quiet_NaN ());
          Q_local.fill (std::numeric_limits<Scalar>::quiet_NaN ());
      }
      Matrix<Ordinal, Scalar> R (ncols, ncols, Scalar(0));

      // Generate the test problem.
      distributedTestProblem (generator, A_local, ordinalComm.get(), scalarComm.get());
      if (b_debug) {
        scalarComm->barrier ();
        if (my_rank == 0) {
          cerr << "-- Generated test problem." << endl;
        }
      }

      // Make sure that the test problem (the matrix to factor) was
      // distributed correctly.
      if (b_extra_debug && b_debug) {
        if (my_rank == 0) {
          cerr << "Test matrix A:" << endl;
        }
        scalarComm->barrier ();
        printGlobalMatrix (cerr, A_local, scalarComm.get(), ordinalComm.get());
        scalarComm->barrier ();
      }

      // Factoring the matrix stored in A_local overwrites it, so we
      // make a copy of A_local.  Initialize with NaNs to make sure
      // that cache blocking works correctly (if applicable).
      Matrix< Ordinal, Scalar > A_copy (nrows_local, ncols);
      if (std::numeric_limits< Scalar >::has_quiet_NaN) {
        A_copy.fill (std::numeric_limits< Scalar >::quiet_NaN ());
      }

      // actual_cache_size_hint: "cache_size_hint" is just a
      // suggestion.  TSQR determines the cache size hint itself;
      // this remembers it so we can print it out later.
      size_t actual_cache_size_hint;

      if (which == "MpiTbbTSQR") {
#ifdef HAVE_KOKKOSTSQR_TBB
        using Teuchos::RCP;
        typedef TSQR::TBB::TbbTsqr< Ordinal, Scalar > node_tsqr_type;
        typedef TSQR::DistTsqr< Ordinal, Scalar > dist_tsqr_type;
        typedef Tsqr< Ordinal, Scalar, node_tsqr_type, dist_tsqr_type > tsqr_type;

        RCP< node_tsqr_type > node_tsqr (new node_tsqr_type (num_cores, cache_size_hint));
        RCP< dist_tsqr_type > dist_tsqr (new dist_tsqr_type (scalarComm));
        tsqr_type tsqr (node_tsqr, dist_tsqr);

        // Compute the factorization and explicit Q factor.
        TsqrVerifier< tsqr_type >::verify (tsqr, scalarComm, A_local, A_copy,
                                           Q_local, R, contiguousCacheBlocks,
                                           b_debug);
        // Save the "actual" cache block size
        actual_cache_size_hint = tsqr.cache_size_hint();
#else
        throw std::logic_error("TSQR not built with Intel TBB support");
#endif // HAVE_KOKKOSTSQR_TBB
      }
      else if (which == "MpiSeqTSQR") {
        using Teuchos::RCP;
        typedef SequentialTsqr< Ordinal, Scalar > node_tsqr_type;
        typedef TSQR::DistTsqr< Ordinal, Scalar > dist_tsqr_type;
        typedef Tsqr< Ordinal, Scalar, node_tsqr_type, dist_tsqr_type > tsqr_type;

        RCP< node_tsqr_type > node_tsqr (new node_tsqr_type (cache_size_hint));
        RCP< dist_tsqr_type > dist_tsqr (new dist_tsqr_type (scalarComm));
        tsqr_type tsqr (node_tsqr, dist_tsqr);

        // Compute the factorization and explicit Q factor.
        TsqrVerifier< tsqr_type >::verify (tsqr, scalarComm, A_local, A_copy,
                                           Q_local, R, contiguousCacheBlocks,
                                           b_debug);
        // Save the "actual" cache block size
        actual_cache_size_hint = tsqr.cache_size_hint();
      }
      else {
        throw std::logic_error("Unknown TSQR implementation type \"" + which + "\"");
      }

      // Print out the Q and R factors
      if (b_extra_debug && b_debug) {
        if (my_rank == 0) {
          cerr << endl << "Q factor:" << endl;
        }
        scalarComm->barrier ();
        printGlobalMatrix (cerr, Q_local, scalarComm.get (), ordinalComm.get ());
        scalarComm->barrier ();
        if (my_rank == 0) {
          cerr << endl << "R factor:" << endl;
          print_local_matrix (cerr, ncols, ncols, R.get(), R.lda());
          cerr << endl;
        }
        scalarComm->barrier ();
      }

      // Test accuracy of the resulting factorization
      std::vector< magnitude_type > results =
        global_verify (nrows_local, ncols, A_local.get(), A_local.lda(),
                       Q_local.get(), Q_local.lda(), R.get(), R.lda(),
                       scalarComm.get());
      if (b_debug) {
        scalarComm->barrier ();
        if (my_rank == 0) {
          cerr << "-- Finished global_verify" << endl;
        }
      }

      // Print the results on Proc 0.
      if (my_rank == 0) {
        if (human_readable) {
          std::string human_readable_name;

          if (which == "MpiSeqTSQR") {
            human_readable_name = "MPI parallel / cache-blocked TSQR";
          }
          else if (which == "MpiTbbTSQR") {
#ifdef HAVE_KOKKOSTSQR_TBB
            human_readable_name = "MPI parallel / TBB parallel / cache-blocked TSQR";
#else
            throw std::logic_error("TSQR not built with Intel TBB support");
#endif // HAVE_KOKKOSTSQR_TBB
          }
          else {
            throw std::logic_error("Unknown TSQR implementation type \"" + which + "\"");
          }

          cout << human_readable_name << ":" << endl
               << "Scalar type: " << scalarTypeName << endl
               << "# rows: " << nrows_global << endl
               << "# columns: " << ncols << endl
               << "# MPI processes: " << nprocs << endl;
#ifdef HAVE_KOKKOSTSQR_TBB
          if (which == "MpiTbbTSQR")
            cout << "# cores per process = " << num_cores << endl;
#endif // HAVE_KOKKOSTSQR_TBB
          cout << "Cache size hint in bytes: " << actual_cache_size_hint << endl
               << "Contiguous cache blocks? " << contiguousCacheBlocks << endl
               << "Absolute residual $\\| A - Q R \\|_2: "
               << results[0] << endl
               << "Absolute orthogonality $\\| I - Q^* Q \\|_2$: "
               << results[1] << endl
               << "Test matrix norm $\\| A \\|_F$: "
               << results[2] << endl
               << endl;
        }
        else {
          if (printFieldNames) {
            cout << "%"
                 << "method"
                 << ",scalarType"
                 << ",globalNumRows"
                 << ",numCols"
                 << ",numProcs"
                 << ",numCores"
                 << ",cacheSizeHint"
                 << ",contiguousCacheBlocks"
                 << ",absFrobResid"
                 << ",absFrobOrthog"
                 << ",frobA" << endl;
          }

          cout << which
               << "," << scalarTypeName
               << "," << nrows_global
               << "," << ncols
               << "," << nprocs;
#ifdef HAVE_KOKKOSTSQR_TBB
          if (which == "MpiTbbTSQR") {
            cout << "," << num_cores;
          } else {
            cout << ",1";
          }
#else
          cout << ",1" << endl;
#endif // HAVE_KOKKOSTSQR_TBB
          cout << "," << actual_cache_size_hint
               << "," << contiguousCacheBlocks
               << "," << results[0]
               << "," << results[1]
               << "," << results[2]
               << endl;
        }
      }
    }


    template<class TsqrBase, class TimerType>
    double
    do_tsqr_benchmark (const std::string& which,
                       TsqrBase& tsqr,
                       const Teuchos::RCP< MessengerBase< typename TsqrBase::scalar_type > >& messenger,
                       const Matrix< typename TsqrBase::ordinal_type, typename TsqrBase::scalar_type >& A_local,
                       Matrix< typename TsqrBase::ordinal_type, typename TsqrBase::scalar_type >& A_copy,
                       Matrix< typename TsqrBase::ordinal_type, typename TsqrBase::scalar_type >& Q_local,
                       Matrix< typename TsqrBase::ordinal_type, typename TsqrBase::scalar_type >& R,
                       const int ntrials,
                       const bool contiguousCacheBlocks,
                       const bool human_readable,
                       const bool b_debug = false)
    {
      typedef typename TsqrBase::FactorOutput factor_output_type;
      typedef typename TsqrBase::ordinal_type ordinal_type;
      using std::cerr;
      using std::cout;
      using std::endl;

      const ordinal_type nrows_local = A_local.nrows();
      const ordinal_type ncols = A_local.ncols();

      if (contiguousCacheBlocks) {
        tsqr.cache_block (nrows_local, ncols, A_copy.get(),
                          A_local.get(), A_local.lda());
        if (b_debug) {
          messenger->barrier ();
          if (messenger->rank () == 0) {
            cerr << "-- Cache-blocked input matrix to factor." << endl;
          }
        }
      }
      else {
        deep_copy (A_copy, A_local);
      }

      if (b_debug) {
        messenger->barrier ();
        if (messenger->rank () == 0) {
          cerr << "-- Starting timing loop" << endl;
        }
      }

      // Benchmark TSQR for ntrials trials.  The answer (the numerical
      // results of the factorization) is only valid if ntrials == 1,
      // but this is a benchmark and not a verification routine.  Call
      // tsqr_verify() if you want to determine whether TSQR computes
      // the right answer.
      //
      // Name of timer doesn't matter here; we only need the timing.
      TSQR::Test::verifyTimerConcept< TimerType >();
      TimerType timer (which);


      const bool testFactorExplicit = true;
      double tsqr_timing;
      if (testFactorExplicit)
        {
          timer.start();
          for (int trial_num = 0; trial_num < ntrials; ++trial_num)
            tsqr.factorExplicit (A_copy.view(), Q_local.view(), R.view(),
                                 contiguousCacheBlocks);
          tsqr_timing = timer.stop();
        }
      else
        {
          timer.start();
          for (int trial_num = 0; trial_num < ntrials; ++trial_num)
            {
              // Factor the matrix and compute the explicit Q factor.
              // Don't worry about the fact that we're overwriting the
              // input; this is a benchmark, not a numerical verification
              // test.  (We have the latter implemented as tsqr_verify()
              // in this file.)  For the same reason, don't worry about
              // un-cache-blocking the output (when cache blocks are
              // stored contiguously).
              factor_output_type factor_output =
                tsqr.factor (nrows_local, ncols, A_copy.get(), A_copy.lda(),
                             R.get(), R.lda(), contiguousCacheBlocks);
              tsqr.explicit_Q (nrows_local,
                               ncols, A_copy.get(), A_copy.lda(), factor_output,
                               ncols, Q_local.get(), Q_local.lda(),
                               contiguousCacheBlocks);
              // Timings in debug mode likely won't make sense, because
              // Proc 0 is outputting the debug messages to cerr.
              // Nevertheless, we don't put any "if(b_debug)" calls in the
              // timing loop.
            }
          // Compute the resulting total time (in seconds) to execute
          // ntrials runs of Tsqr::factor() and Tsqr::explicit_Q().  The
          // time may differ on different MPI processes.
          tsqr_timing = timer.stop();
        }

      if (b_debug)
        {
          messenger->barrier();
          if (messenger->rank() == 0)
            cerr << "-- Finished timing loop" << endl;
        }
      return tsqr_timing;
    }

    /// \function benchmarkTsqr
    /// \brief Benchmark parallel TSQR and report timings to stdout
    ///
    /// Benchmark the MPI-parallel TSQR implementation specified by
    /// the "which" parameter (either with cache-blocked TSQR or
    /// TBB-parallel cache-blocked TSQR as the node-level
    /// implementation), for "ntrials" trials.  Print the stdout the
    /// cumulative run time (in seconds) for all ntrials trials.
    ///
    /// \param which [in] Valid values: "MpiTbbTSQR" (for TBB-parallel
    ///   node-level TSQR underneath MPI-parallel TSQR), "MpiSeqTSQR"
    ///   (for cache-blocked sequential node-level TSQR underneath
    ///   MPI-parallel TSQR)
    ///
    /// \param scalarTypeName [in] Name of the Scalar type
    ///
    /// \param generator [in/out] Normal(0,1) (pseudo)random number
    ///   generator.  Only touched on MPI process 0.  Used to generate
    ///   random test matrices for the factorization.
    ///
    /// \param ntrials [in] Number of trials to use in the benchmark.
    ///   Reported timings are cumulative over all trials.
    ///
    /// \param nrows_global [in] Number of rows in the entire test
    ///   matrix (over all processes) to generate.  The matrix will be
    ///   divided up in blocks of contiguous rows among the processes.
    ///
    /// \param ncols [in] Number of columns in the test matrix to
    ///   generate.
    ///
    /// \param ordinalComm [in/out] Object for communicating Ordinal
    ///   (integer index) objects among the processes
    ///
    /// \param scalarComm [in/out] Object for communicating Scalar
    ///   (matrix data) objects among the processes
    ///
    /// \param num_cores [in] Number of cores to use per MPI process
    ///   for Intel TBB parallelism within that process
    ///
    /// \param cache_size_hint [in] Cache block size (per core) in
    ///   bytes.  If zero, a sensible default is used.
    ///
    /// \param contiguousCacheBlocks [in] Whether cache blocks
    ///   should be stored contiguously
    ///
    /// \param printFieldNames [in] Whether to print field names (only
    ///   appliable if not human_readable)
    ///
    /// \param human_readable [in] Whether output should be human
    ///   readable, or machine parseable
    ///
    /// \param b_debug [in] Whether to print debug output
    ///
    template<class Ordinal, class Scalar, class Generator, class TimerType>
    void
    benchmarkTsqr (const std::string& which,
                   const std::string& scalarTypeName,
                   Generator& generator,
                   const int ntrials,
                   const Ordinal nrows_global,
                   const Ordinal ncols,
                   const Teuchos::RCP< MessengerBase< Ordinal > >& ordinalComm,
                   const Teuchos::RCP< MessengerBase< Scalar > >& scalarComm,
                   const Ordinal num_cores,
                   const size_t cache_size_hint,
                   const bool contiguousCacheBlocks,
                   const bool printFieldNames,
                   const bool human_readable,
                   const bool b_debug)
    {
      using std::cerr;
      using std::cout;
      using std::endl;

      TSQR::Test::verifyTimerConcept< TimerType >();
      const bool b_extra_debug = false;
      const int nprocs = scalarComm->size();
      const int my_rank = scalarComm->rank();
      if (b_debug)
        {
          scalarComm->barrier();
          if (my_rank == 0)
            cerr << "tsqr_benchmark:" << endl;
          scalarComm->barrier();
        }
      const Ordinal nrows_local = numLocalRows (nrows_global, my_rank, nprocs);

      // Set up storage for the test problem.
      Matrix< Ordinal, Scalar > A_local (nrows_local, ncols);
      Matrix< Ordinal, Scalar > Q_local (nrows_local, ncols);
      if (std::numeric_limits< Scalar >::has_quiet_NaN)
        {
          A_local.fill (std::numeric_limits< Scalar >::quiet_NaN());
          Q_local.fill (std::numeric_limits< Scalar >::quiet_NaN());
        }
      Matrix< Ordinal, Scalar > R (ncols, ncols, Scalar(0));

      // Generate the test problem.
      distributedTestProblem (generator, A_local, ordinalComm.get(), scalarComm.get());
      if (b_debug)
        {
          scalarComm->barrier();
          if (my_rank == 0)
            cerr << "-- Generated test problem." << endl;
        }

      // Make sure that the test problem (the matrix to factor) was
      // distributed correctly.
      if (b_extra_debug && b_debug)
        {
          if (my_rank == 0)
            cerr << "Test matrix A:" << endl;
          scalarComm->barrier ();
          printGlobalMatrix (cerr, A_local, scalarComm.get(), ordinalComm.get());
          scalarComm->barrier ();
        }

      // Factoring the matrix stored in A_local overwrites it, so we
      // make a copy of A_local.  If specified, rearrange cache blocks
      // in the copy.  Initialize with NaNs to make sure that cache
      // blocking worked correctly.
      Matrix< Ordinal, Scalar > A_copy (nrows_local, ncols);
      if (std::numeric_limits< Scalar >::has_quiet_NaN)
        A_copy.fill (std::numeric_limits< Scalar >::quiet_NaN());

      // actual_cache_size_hint: "cache_size_hint" is just a
      // suggestion.  TSQR determines the cache block size itself;
      // this remembers it so we can print it out later.
      size_t actual_cache_size_hint;
      // Run time (in seconds, as a double-precision floating-point
      // value) for TSQR on this MPI node.
      double tsqr_timing;

      if (which == "MpiTbbTSQR")
        {
#ifdef HAVE_KOKKOSTSQR_TBB
          using Teuchos::RCP;
          typedef TSQR::TBB::TbbTsqr< Ordinal, Scalar > node_tsqr_type;
          typedef TSQR::DistTsqr< Ordinal, Scalar > dist_tsqr_type;
          typedef Tsqr< Ordinal, Scalar, node_tsqr_type, dist_tsqr_type > tsqr_type;

          RCP< node_tsqr_type > nodeTsqr (new node_tsqr_type (num_cores, cache_size_hint));
          RCP< dist_tsqr_type > distTsqr (new dist_tsqr_type (scalarComm));
          tsqr_type tsqr (nodeTsqr, distTsqr);

          // Run the benchmark.
          tsqr_timing =
            do_tsqr_benchmark< tsqr_type, TimerType > (which, tsqr, scalarComm, A_local,
                                                       A_copy, Q_local, R, ntrials,
                                                       contiguousCacheBlocks,
                                                       human_readable, b_debug);

          // Save the "actual" cache block size
          actual_cache_size_hint = tsqr.cache_size_hint();
#else
          throw std::logic_error("TSQR not built with Intel TBB support");
#endif // HAVE_KOKKOSTSQR_TBB
        }
      else if (which == "MpiSeqTSQR")
        {
          using Teuchos::RCP;
          typedef SequentialTsqr< Ordinal, Scalar > node_tsqr_type;
          typedef TSQR::DistTsqr< Ordinal, Scalar > dist_tsqr_type;
          typedef Tsqr< Ordinal, Scalar, node_tsqr_type, dist_tsqr_type > tsqr_type;

          // Set up TSQR.
          RCP< node_tsqr_type > nodeTsqr (new node_tsqr_type (cache_size_hint));
          RCP< dist_tsqr_type > distTsqr (new dist_tsqr_type (scalarComm));
          tsqr_type tsqr (nodeTsqr, distTsqr);

          // Run the benchmark.
          tsqr_timing =
            do_tsqr_benchmark< tsqr_type, TimerType > (which, tsqr, scalarComm, A_local,
                                                       A_copy, Q_local, R, ntrials,
                                                       contiguousCacheBlocks,
                                                       human_readable, b_debug);
          // Save the "actual" cache block size
          actual_cache_size_hint = tsqr.cache_size_hint();
        }
      else
        throw std::logic_error("Unknown TSQR implementation type \"" + which + "\"");

      // Find the min and max TSQR timing on all processors.
      const double min_tsqr_timing = scalarComm->globalMin (tsqr_timing);
      const double max_tsqr_timing = scalarComm->globalMax (tsqr_timing);

      // Print the results on Proc 0.
      if (my_rank == 0)
        {
          if (human_readable)
            {
              std::string human_readable_name;

              if (which == "MpiSeqTSQR")
                human_readable_name = "MPI parallel / cache-blocked TSQR";
              else if (which == "MpiTbbTSQR")
                {
#ifdef HAVE_KOKKOSTSQR_TBB
                  human_readable_name = "MPI parallel / TBB parallel / cache-blocked TSQR";
#else
                  throw std::logic_error("TSQR not built with Intel TBB support");
#endif // HAVE_KOKKOSTSQR_TBB
                }
              else
                throw std::logic_error("Unknown TSQR implementation type \"" + which + "\"");

              cout << human_readable_name << ":" << endl
                   << "Scalar type: " << scalarTypeName << endl
                   << "# rows: " << nrows_global << endl
                   << "# columns: " << ncols << endl
                   << "# MPI processes: " << nprocs << endl;

#ifdef HAVE_KOKKOSTSQR_TBB
              if (which == "MpiTbbTSQR")
                cout << "# cores per process: " << num_cores << endl;
#endif // HAVE_KOKKOSTSQR_TBB

              cout << "Cache size hint in bytes: " << actual_cache_size_hint << endl
                   << "contiguous cache blocks? " << contiguousCacheBlocks << endl
                   << "# trials: " << ntrials << endl
                   << "Min total time (s) over all MPI processes: "
                   << min_tsqr_timing << endl
                   << "Max total time (s) over all MPI processes: "
                   << max_tsqr_timing << endl
                   << endl;
            }
          else
            {
              if (printFieldNames)
                {
                  cout << "%"
                       << "method"
                       << ",scalarType"
                       << ",globalNumRows"
                       << ",numCols"
                       << ",numProcs"
                       << ",numCores"
                       << ",cacheSizeHint"
                       << ",contiguousCacheBlocks"
                       << ",numTrials"
                       << ",minTiming"
                       << ",maxTiming"
                       << endl;
                }
              cout << which
                   << "," << scalarTypeName
                   << "," << nrows_global
                   << "," << ncols
                   << "," << nprocs;
#ifdef HAVE_KOKKOSTSQR_TBB
              if (which == "MpiTbbTSQR")
                cout << "," << num_cores;
              else
                cout << ",1";
#else
              cout << ",1";
#endif // HAVE_KOKKOSTSQR_TBB
              cout << "," << actual_cache_size_hint
                   << "," << contiguousCacheBlocks
                   << "," << ntrials
                   << "," << min_tsqr_timing
                   << "," << max_tsqr_timing
                   << endl;
            }
        }
    }


  } // namespace Test
} // namespace TSQR

#endif // __TSQR_Test_TsqrTest_hpp