Intrepid2_TensorViewIterator.hpp

Go to the documentation of this file.

// @HEADER
// ************************************************************************
//
//                           Intrepid2 Package
//                 Copyright (2007) Sandia Corporation
//
// Under terms of Contract DE-AC04-94AL85000, there is a non-exclusive
// license for use of this work by or on behalf of the U.S. Government.
//
// 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 Kyungjoo Kim  (kyukim@sandia.gov),
//                    Mauro Perego  (mperego@sandia.gov), or
//                    Nate Roberts  (nvrober@sandia.gov)
//
// ************************************************************************
// @HEADER

#ifndef Intrepid2_TensorViewIterator_h
#define Intrepid2_TensorViewIterator_h

#include "Intrepid2_DeviceAssert.hpp"
#include "Intrepid2_ViewIterator.hpp"

#include <Kokkos_Core.hpp>

#include <vector>

namespace Intrepid2
{
  template<class TensorViewType, class ViewType1, class ViewType2 ,typename ScalarType>
  class TensorViewIterator
  {
  public:
    enum RankCombinationType : int
    {
      DIMENSION_MATCH,
      TENSOR_PRODUCT,
      TENSOR_CONTRACTION
    };
    using RankCombinationViewType = Kokkos::View<RankCombinationType*, typename TensorViewType::device_type>;
  protected:
    
    ViewIterator<TensorViewType, ScalarType> tensor_view_iterator_;
    ViewIterator<ViewType1, ScalarType> view1_iterator_;
    ViewIterator<ViewType2, ScalarType> view2_iterator_;
    
    RankCombinationViewType rank_combination_types_;
  public:
    KOKKOS_INLINE_FUNCTION
    TensorViewIterator(TensorViewType tensor_view, ViewType1 view1, ViewType2 view2,
        RankCombinationViewType rank_combination_types)
    :
    tensor_view_iterator_(tensor_view),
    view1_iterator_(view1),
    view2_iterator_(view2),
    rank_combination_types_(rank_combination_types)
    {
      // rank_combination_type should have length equal to the maximum rank of the views provided
      /*
       Examples:
       1. vector dot product in third dimension: {DIMENSION_MATCH, DIMENSION_MATCH, TENSOR_CONTRACTION}
       - view1 and view2 should both be rank 3, and should match in all dimensions
       - tensor_view should be rank 2, and should match view1 and view2 in first two dimensions
       2. vector outer product in third dimension: {DIMENSION_MATCH, DIMENSION_MATCH, TENSOR_PRODUCT}
       - view1 and view2 should both be rank 3, and should match in first two dimensions
       - tensor_view should be rank 3, and should match view1 and view2 in first two dimensions
       - in third dimension, tensor_view should have dimension equal to the product of the third dimension of view1 and the third dimension of view2
       3. rank-3 view1 treated as vector times scalar rank-2 view2: {DIMENSION_MATCH, DIMENSION_MATCH, TENSOR_PRODUCT}
       - here, the rank-2 view2 is interpreted as having an extent 1 third dimension
       
       We only allow TENSOR_CONTRACTION in final dimension(s)
       */
      // check that the above rules are satisfied:
      unsigned max_component_rank = (view1.rank() > view2.rank()) ? view1.rank() : view2.rank();
      unsigned max_rank           = (tensor_view.rank() > max_component_rank) ? tensor_view.rank() : max_component_rank;
      
      INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(rank_combination_types.extent(0) != max_rank, std::invalid_argument, "need to provide RankCombinationType for the largest-rank View");
      
      unsigned expected_rank = 0;
      bool contracting = false;
      for (unsigned d=0; d<rank_combination_types.extent(0); d++)
      {
        if (rank_combination_types[d] == TENSOR_CONTRACTION)
        {
          // check that view1 and view2 agree on the length of this dimension
          INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(view1.extent_int(d) != view2.extent_int(d), std::invalid_argument, "Contractions can only occur along ranks of equal length");
          contracting = true;
        }
        else
        {
          if (!contracting)
          {
            expected_rank++;
            if (rank_combination_types[d] == TENSOR_PRODUCT)
            {
              INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(tensor_view.extent_int(d) != view1.extent_int(d) * view2.extent_int(d), std::invalid_argument, "For TENSOR_PRODUCT rank combination, the tensor View must have length in that dimension equal to the product of the two component views in that dimension");
            }
            else // matching
            {
              INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(view1.extent_int(d) != view2.extent_int(d), std::invalid_argument, "For DIMENSION_MATCH rank combination, all three views must have length equal to each other in that rank");
              INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(tensor_view.extent_int(d) != view1.extent_int(d), std::invalid_argument, "For DIMENSION_MATCH rank combination, all three views must have length equal to each other in that rank");
            }
          }
          else
          {
            INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(contracting, std::invalid_argument, "encountered a non-contraction rank combination after a contraction; contractions can only go at the end");
          }
        }
      }
      INTREPID2_TEST_FOR_EXCEPTION_DEVICE_SAFE(expected_rank != tensor_view.rank(), std::invalid_argument, "Tensor view does not match expected rank");
    }
    
    KOKKOS_INLINE_FUNCTION
    int nextIncrementRank()
    {
      int view2_next_increment_rank = view2_iterator_.nextIncrementRank();
      int view1_next_increment_rank = view1_iterator_.nextIncrementRank();
      if (view2_next_increment_rank > view1_next_increment_rank) return view2_next_increment_rank;
      else return view1_next_increment_rank;
    }
    
    KOKKOS_INLINE_FUNCTION
    int increment()
    {
      // proceed to the next view1/view2 combination
      // where we're doing a dimension match, then all three iterators should increment in tandem
      // where we're doing a contraction, view1/view2 should increment in tandem, while tensor_view should be fixed
      // where we're doing a tensor product, view1 and tensor_view increment in tandem, while view2 is fixed
      
      // note that regardless of the choice, view1 should be incremented, with one exception:
      //   If we are doing a tensor product, then view1 can be understood to be in an interior for loop, and it should loop around.
      //   We can detect this by checking which the least rank that would be updated -- if view2's least rank exceeds view1's, then:
      //   - view1 should be reset, AND
      //   - view2 should be incremented (as should the tensor view)
      int view2_next_increment_rank = view2_iterator_.nextIncrementRank();
      int view1_next_increment_rank = view1_iterator_.nextIncrementRank();
      if (view2_next_increment_rank > view1_next_increment_rank)
      {
        // if we get here, we should be doing a tensor product in the view2 rank that will change
        device_assert(rank_combination_types_[view2_next_increment_rank]==TENSOR_PRODUCT);
        view1_iterator_.reset(view2_next_increment_rank); // set to 0 from the tensor product rank inward -- this is "looping around"
        view2_iterator_.increment();
        tensor_view_iterator_.increment();
        return view2_next_increment_rank;
      }
      else
      {
        int view1_rank_change = view1_iterator_.increment();
        if (view1_rank_change >= 0)
        {
          switch (rank_combination_types_[view1_rank_change])
          {
            case DIMENSION_MATCH:
              view2_iterator_.increment();
              tensor_view_iterator_.increment();
              break;
            case TENSOR_PRODUCT:
              // view1 increments fastest; the only time we increment view2 is when view1 loops around; we handle that above
              tensor_view_iterator_.increment();
              break;
            case TENSOR_CONTRACTION:
              // view1 and view2 increment in tandem; we don't increment tensor_view while contraction is taking place
              view2_iterator_.increment();
          }
        }
        return view1_rank_change;
      }
    }
    
    KOKKOS_INLINE_FUNCTION
    void setLocation(const Kokkos::Array<int,7> location)
    {
      view1_iterator_.setLocation(location);
      view2_iterator_.setLocation(location);
      tensor_view_iterator_.setLocation(location);
    }
    
    KOKKOS_INLINE_FUNCTION
    void setLocation(Kokkos::Array<int,7> location1, Kokkos::Array<int,7> location2)
    {
      view1_iterator_.setLocation(location1);
      view2_iterator_.setLocation(location2);
      Kokkos::Array<int,7> tensor_location = location1;
      for (unsigned d=0; d<rank_combination_types_.extent(0); d++)
      {
        switch (rank_combination_types_[d])
        {
          case TENSOR_PRODUCT:
            // view1 index is fastest-moving:
            tensor_location[d] = location2[d] * view1_iterator_.getExtent(d) + location1[d];
            break;
          case DIMENSION_MATCH:
            // we copied location1 into tensor_location to initialize -- that's correct in this dimension
            break;
          case TENSOR_CONTRACTION:
            tensor_location[d] = 0;
            break;
        }
      }
#ifdef HAVE_INTREPID2_DEBUG
      // check that the location makes sense
      for (unsigned d=0; d<rank_combination_types_.extent(0); d++)
      {
        switch (rank_combination_types_[d])
        {
          case TENSOR_PRODUCT:
            // in this case, the two indices are independent
            break;
          case DIMENSION_MATCH:
          case TENSOR_CONTRACTION:
            device_assert(location1[d] == location2[d]);
            break;
        }
        // let's check that the indices are in bounds:
        device_assert(location1[d] < view1_iterator_.getExtent(d));
        device_assert(location2[d] < view2_iterator_.getExtent(d));
        device_assert(tensor_location[d] < tensor_view_iterator_.getExtent(d));
      }
#endif
      tensor_view_iterator_.setLocation(tensor_location);
    }
    
    KOKKOS_INLINE_FUNCTION
    ScalarType getView1Entry()
    {
      return view1_iterator_.get();
    }
    
    KOKKOS_INLINE_FUNCTION
    ScalarType getView2Entry()
    {
      return view2_iterator_.get();
    }
    
    KOKKOS_INLINE_FUNCTION
    void set(ScalarType value)
    {
      tensor_view_iterator_.set(value);
    }
  };

} // namespace Intrepid2

#endif /* Intrepid2_TensorViewIterator_h */