|
Intrepid2
|
Implementation of a general sum factorization algorithm, using a novel approach developed by Roberts, for integration. Uses hierarchical parallelism. More...
#include <Intrepid2_IntegrationToolsDef.hpp>
Public Member Functions | |
| F_IntegratePointValueCache (Data< Scalar, DeviceType > integralData, TensorData< Scalar, DeviceType > leftComponent, Data< Scalar, DeviceType > composedTransform, TensorData< Scalar, DeviceType > rightComponent, TensorData< Scalar, DeviceType > cellMeasures, int a_offset, int b_offset, int leftFieldOrdinalOffset, int rightFieldOrdinalOffset) | |
| template<size_t maxComponents, size_t numComponents = maxComponents> | |
| KOKKOS_INLINE_FUNCTION int | incrementArgument (Kokkos::Array< int, maxComponents > &arguments, const Kokkos::Array< int, maxComponents > &bounds) const |
| KOKKOS_INLINE_FUNCTION int | incrementArgument (Kokkos::Array< int, Parameters::MaxTensorComponents > &arguments, const Kokkos::Array< int, Parameters::MaxTensorComponents > &bounds, const int &numComponents) const |
| runtime-sized variant of incrementArgument; gets used by approximate flop count. | |
| template<size_t maxComponents, size_t numComponents = maxComponents> | |
| KOKKOS_INLINE_FUNCTION int | nextIncrementResult (const Kokkos::Array< int, maxComponents > &arguments, const Kokkos::Array< int, maxComponents > &bounds) const |
| KOKKOS_INLINE_FUNCTION int | nextIncrementResult (const Kokkos::Array< int, Parameters::MaxTensorComponents > &arguments, const Kokkos::Array< int, Parameters::MaxTensorComponents > &bounds, const int &numComponents) const |
| runtime-sized variant of nextIncrementResult; gets used by approximate flop count. | |
| template<size_t maxComponents, size_t numComponents = maxComponents> | |
| KOKKOS_INLINE_FUNCTION int | relativeEnumerationIndex (const Kokkos::Array< int, maxComponents > &arguments, const Kokkos::Array< int, maxComponents > &bounds, const int startIndex) const |
| template<int rank> | |
| KOKKOS_INLINE_FUNCTION enable_if_t< rank==3 &&rank==integralViewRank, Scalar & > | integralViewEntry (const IntegralViewType &integralView, const int &cellDataOrdinal, const int &i, const int &j) const |
| template<int rank> | |
| KOKKOS_INLINE_FUNCTION enable_if_t< rank==2 &&rank==integralViewRank, Scalar & > | integralViewEntry (const IntegralViewType &integralView, const int &cellDataOrdinal, const int &i, const int &j) const |
| KOKKOS_INLINE_FUNCTION void | runSpecialized3 (const TeamMember &teamMember) const |
| Hand-coded 3-component version. | |
| template<size_t numTensorComponents> | |
| KOKKOS_INLINE_FUNCTION void | run (const TeamMember &teamMember) const |
| KOKKOS_INLINE_FUNCTION void | operator() (const TeamMember &teamMember) const |
| long | approximateFlopCountPerCell () const |
| returns an estimate of the number of floating point operations per cell (counting sums, subtractions, divisions, and multiplies, each of which counts as one operation). | |
| int | teamSize (const int &maxTeamSizeFromKokkos) const |
| returns the team size that should be provided to the policy constructor, based on the Kokkos maximum and the amount of thread parallelism we have available. | |
| size_t | team_shmem_size (int numThreads) const |
| Provide the shared memory capacity. | |
Private Types | |
| using | ExecutionSpace = typename DeviceType::execution_space |
| using | TeamPolicy = Kokkos::TeamPolicy< DeviceType > |
| using | TeamMember = typename TeamPolicy::member_type |
| using | IntegralViewType = Kokkos::View< typename RankExpander< Scalar, integralViewRank >::value_type, DeviceType > |
Private Attributes | |
| IntegralViewType | integralView_ |
| TensorData< Scalar, DeviceType > | leftComponent_ |
| Data< Scalar, DeviceType > | composedTransform_ |
| TensorData< Scalar, DeviceType > | rightComponent_ |
| TensorData< Scalar, DeviceType > | cellMeasures_ |
| int | a_offset_ |
| int | b_offset_ |
| int | leftComponentSpan_ |
| int | rightComponentSpan_ |
| int | numTensorComponents_ |
| int | leftFieldOrdinalOffset_ |
| int | rightFieldOrdinalOffset_ |
| size_t | fad_size_output_ = 0 |
|
Kokkos::Array< int, Parameters::MaxTensorComponents > | leftFieldBounds_ |
|
Kokkos::Array< int, Parameters::MaxTensorComponents > | rightFieldBounds_ |
|
Kokkos::Array< int, Parameters::MaxTensorComponents > | pointBounds_ |
| int | maxFieldsLeft_ |
| int | maxFieldsRight_ |
| int | maxPointCount_ |
Implementation of a general sum factorization algorithm, using a novel approach developed by Roberts, for integration. Uses hierarchical parallelism.
Whereas F_Integrate, and Mora and Demkowicz, and all others we are aware of, cache partial sums at intermediate component levels — the cached values are indexed by component basis ordinals — we integrate the first component in its dimension(s) and store values for integration points in the remaining dimensions, so that our caches are indexed by point ordinals. If there are L_x, L_y, and L_z quadrature points in dimensions x,y,z, we require a cache of size L_y * L_z +1 for a 3D, 3-component integral. The standard approach requires a cache of size (p_x+1)*(p_y+1). So long as one is not over-integrating by too much, these sizes are about the same. The real advantage of our approach here is (we expect) that it improves data locality.
Definition at line 992 of file Intrepid2_IntegrationToolsDef.hpp.
1.8.5