RTIExample.cpp

Go to the documentation of this file.

/*
// @HEADER
// ***********************************************************************
// 
//          Tpetra: Templated Linear Algebra Services Package
//                 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
*/

#include <Teuchos_GlobalMPISession.hpp>
#include <Teuchos_oblackholestream.hpp>
#include <Teuchos_CommandLineProcessor.hpp>
#include <Teuchos_TimeMonitor.hpp>

#include <Kokkos_DefaultNode.hpp>

#include "Tpetra_Version.hpp"
#include "Tpetra_DefaultPlatform.hpp"
#include "Tpetra_MultiVector.hpp"
#include "Tpetra_RTI.hpp"

#include <iostream>
#include <iomanip>
#include <functional>
#include <utility>

namespace TpetraExamples {

  template <class Arg1, class Arg2, class MultPrec>
  class mprec_mult : public std::binary_function<Arg1,Arg2,MultPrec> {
  public:
    mprec_mult() {}
    inline MultPrec 
    operator() (const Arg1& x, const Arg2& y) const {
      return (MultPrec)(x) * (MultPrec)(y);
    }
  };

  template <class T1, class T2, class Op>
  class pair_op : 
    public std::binary_function<std::pair<T1,T2>, std::pair<T1,T2>, std::pair<T1,T2>> {
  private:
    Op op_;
  public:
    pair_op(Op op) : op_(op) {}
    inline std::pair<T1,T2> 
    operator() (const std::pair<T1,T2>& a, const std::pair<T1,T2>& b) const {
      return std::make_pair (op_ (a.first, b.first), op_ (a.second, b.second));
    }
  };

  template <class T1, class T2, class Op>
  pair_op<T1,T2,Op> make_pair_op (Op op) { 
    return pair_op<T1,T2,Op>(op); 
  }

}

namespace Tpetra {
  namespace RTI {
    //
    // Specialization of ZeroOp for a pair of two different types.
    //
    template <class T1, class T2>
    class ZeroOp<std::pair<T1,T2>> {
      public:
      static inline std::pair<T1,T2> identity() {
        return std::make_pair( Teuchos::ScalarTraits<T1>::zero(), 
                               Teuchos::ScalarTraits<T2>::zero() );
      }
    };
  }
}

int main(int argc, char *argv[])
{
  //
  // Set up MPI, if applicable.
  //
  Teuchos::oblackholestream blackhole;
  Teuchos::GlobalMPISession mpiSession(&argc,&argv,&blackhole);

  using Teuchos::TimeMonitor;
  using Tpetra::RTI::ZeroOp;
  using Tpetra::RTI::reductionGlob;
  using TpetraExamples::mprec_mult;
  using std::make_pair;
  using std::pair;

  // 
  // Get the default communicator and node
  //
  auto &platform = Tpetra::DefaultPlatform::getDefaultPlatform();
  auto comm = platform.getComm();
  auto node = platform.getNode();
  const int myImageID = comm->getRank();

  //
  // Get example parameters from command-line processor
  //  
  bool verbose = (myImageID==0);
  int numGlobal_user = 100*comm->getSize();
  int numTimeTrials = 3;
  Teuchos::CommandLineProcessor cmdp(false,true);
  cmdp.setOption("verbose","quiet",&verbose,"Print messages and results.");
  cmdp.setOption("global-size",&numGlobal_user,"Global test size.");
  cmdp.setOption("num-time-trials",&numTimeTrials,"Number of trials in timing loops.");
  if (cmdp.parse(argc,argv) != Teuchos::CommandLineProcessor::PARSE_SUCCESSFUL) {
    return -1;
  }

  // 
  // Say hello, print some communicator info
  //
  if (verbose) {
    std::cout << "\n" << Tpetra::version() << std::endl;
    std::cout << "Comm info: " << *comm;
    std::cout << "Node type: " << Teuchos::typeName(*node) << std::endl;
    std::cout << std::endl;
  }


  // 
  // Create a simple map with 5 local entries per node
  // 
  Tpetra::global_size_t numGlobalRows = numGlobal_user;
  auto map = Tpetra::createUniformContigMapWithNode<int,int>(numGlobalRows, comm, node);
  const size_t numLocalRows = map->getNodeNumElements();
  auto x = Tpetra::createVector<float>(map),
       y = Tpetra::createVector<float>(map);
  auto z = Tpetra::createVector<double>(map),
       w = Tpetra::createVector<double>(map),
       v = Tpetra::createVector<double>(map);

  //
  // Initialization and simple reduction
  //

  // sets x[i] = 1.0f
  Tpetra::RTI::unary_transform( *x, [](float xi){return 1.0f;} );
  // sets y[i] = x[i]
  Tpetra::RTI::binary_transform( *y, *x, [](float /*yi*/, float xi) {return xi;} );
  // sets y[i] = plus(y[i], x[i]) = y[i] + x[i]
  Tpetra::RTI::binary_transform( *y, *x, std::plus<float>() );
  // compute dot(x,y) in single precision, sum all x[i]*y[i]
  float fresult = Tpetra::RTI::reduce( *x, *y, 
                                       reductionGlob<
                                         ZeroOp<float>>( 
                                         std::multiplies<float>(), 
                                         std::plus<float>()
                                       ) );
  if (verbose) {
    std::cout << std::left << "dot( ones, twos ) result == " << fresult 
        << ", expected value is " << numGlobalRows*2.0f << std::endl;
  }

  //
  // Single precision testing
  //

  // set x = [1, 1e-4, ..., 1e-4]
  {
    Teuchos::ArrayRCP<float> xdata = x->get1dViewNonConst();
    for (size_t i=1; i < numLocalRows; ++i) {
      xdata[i] = 1.0f / 4096.0f;
    }
  }
  // set y[i] = x while computing square of norm2(x) (using y): multiply x[i] * y[i] in float, sum them in float, 0.0f is the additive identity
  fresult = Tpetra::RTI::binary_pre_transform_reduce(*y, *x, reductionGlob<ZeroOp<float>>( [](float /*yi*/, float xi){return xi;}, std::multiplies<float>(), std::plus<float>()) );

  // compute pure float inner product alone, timing it for comparison
  auto timePF = TimeMonitor::getNewTimer("Pure float dot()");
  {
    TimeMonitor lcltimer(*timePF);    
    for (int l=0; l != numTimeTrials; ++l) 
       fresult = Tpetra::RTI::reduce(*x, *y, reductionGlob<ZeroOp<float>>(std::multiplies<float>(), std::plus<float>()) );
  }
  if (verbose) {
    std::cout << std::left << std::endl << std::setw(25) << "pure float result" << " == " << std::setprecision(12) << std::scientific << fresult << std::endl;
  }

  //
  // Mixed precision testing
  // 

  // compute dot(x,y) with double accumulator: multiply x[i] * y[i] in float, sum them in double, 0.0 is the additive identity
  auto timeAD = TimeMonitor::getNewTimer("Double acc. dot()");
  {
    TimeMonitor lcltimer(*timeAD);
    for (int l=0; l != numTimeTrials; ++l) 
       fresult = Tpetra::RTI::reduce(*x, *y, reductionGlob<ZeroOp<double>>(std::multiplies<float>(), std::plus<double>()) );
  }
  if (verbose) {
    std::cout << std::left << std::setw(25) << "double acc. result" << " == " << std::setprecision(12) << std::scientific << fresult << std::endl;
  }

  // compute dot(x,y) in full double precision: multiply x[i] * y[i] in double, sum them in double, 0.0 is the additive identity
  double dresult = 0.0;
  auto timeMAD = TimeMonitor::getNewTimer("Double mult./acc. dot()");
  {
    TimeMonitor lcltimer(*timeMAD);
    for (int l=0; l != numTimeTrials; ++l) 
       dresult = Tpetra::RTI::reduce(*x, *y, reductionGlob< ZeroOp<double>>(
                                                            mprec_mult<float,float,double>(), 
                                                            std::plus<double>()) );
  }
  if (verbose) {
    std::cout << std::left << std::setw(25) << "double mult/add result" << " == " << std::setprecision(12) << std::scientific << dresult << std::endl;
  }

  // compute dot(x,y) in full double precision: multiply x[i] * y[i] in double, sum them in double, 0.0 is the additive identity
  auto timePD = TimeMonitor::getNewTimer("Pure double dot()");
  // set [w,z] = x
  Tpetra::RTI::binary_transform( *w, *x, [](double /*wi*/, float xi) -> double {return xi;} );
  Tpetra::RTI::binary_transform( *z, *x, [](double /*zi*/, float xi) -> double {return xi;} );
  {
    TimeMonitor lcltimer(*timePD);
    for (int l=0; l != numTimeTrials; ++l) 
       dresult = Tpetra::RTI::reduce(*w, *z, reductionGlob< ZeroOp<double>>(
                                                            std::multiplies<double>(), 
                                                            std::plus<double>()) );
  }
  if (verbose) {
    std::cout << std::left << std::setw(25) << "pure double result" << " == " << std::setprecision(12) << std::scientific << dresult << std::endl
              << std::endl;
  }

  //
  // Compute z = alpha*x, where z is double precision and x is single precision
  //
  Tpetra::RTI::binary_transform( *z, *x, [](double /*zi*/, float xi) -> double {return 2.0*xi;} );

  // 
  // Two simultaneous dot products: z'*w and z'*v at the same time, saves a pass through z
  // 
  Tpetra::RTI::unary_transform( *z, [](double xi){return 1.0f;} );
  Tpetra::RTI::unary_transform( *w, [](double xi){return 2.0f;} );
  Tpetra::RTI::unary_transform( *v, [](double xi){return 3.0f;} );
  auto timeTwoDotInd = TimeMonitor::getNewTimer("dot(z,w) and dot(z,v) independently");
  {
    TimeMonitor lcltimer(*timeTwoDotInd);
    for (int l=0; l != numTimeTrials; ++l) 
    {
       dresult = Tpetra::RTI::reduce(*z, *w, reductionGlob< ZeroOp<double>>(
                                                            std::multiplies<double>(), 
                                                            std::plus<double>()) );
       dresult = Tpetra::RTI::reduce(*z, *v, reductionGlob< ZeroOp<double>>(
                                                            std::multiplies<double>(), 
                                                            std::plus<double>()) );
    }
  }
  auto timeTwoDotSim = TimeMonitor::getNewTimer("dot(z,w) and dot(z,v) simultaneously");
  {
    TimeMonitor lcltimer(*timeTwoDotSim);
    pair<double,double> tdres;
    for (int l=0; l != numTimeTrials; ++l) 
    {
      tdres = Tpetra::RTI::reduce(*z, *w, *v, 
                                  reductionGlob<ZeroOp<pair<double,double>>>(
                                                [](double zi, double wi, double vi) { return make_pair(zi*wi, zi*vi); },
                                                TpetraExamples::make_pair_op<double,double>(std::plus<double>()))
                                 );
    }
  }

  //
  // Print timings
  //
  if (verbose) {
    TimeMonitor::summarize( std::cout );
  }

  if (verbose) {
    std::cout << "\nEnd Result: TEST PASSED" << std::endl;
  }
  return 0;
}