HMLP: High-performance Machine Learning Primitives
|
These are data that shared by the whole local tree. Distributed setup inherits mpitree::Setup. More...
#include <gofmm_mpi.hpp>
Public Member Functions | |
void | FromConfiguration (gofmm::Configuration< T > &config, SPDMATRIX &K, SPLITTER &splitter, DistData< STAR, CBLK, pair< T, size_t >> *NN_cblk) |
![]() | |
vector< size_t > | ContainAny (vector< size_t > &queries, size_t target) |
Check if this node contain any query using morton. Notice that queries[] contains gids; thus, morton[] needs to be accessed using gids. More... | |
![]() | |
Configuration (DistanceMetric metric_type, size_t problem_size, size_t leaf_node_size, size_t neighbor_size, size_t maximum_rank, T tolerance, T budget) | |
void | Set (DistanceMetric metric_type, size_t problem_size, size_t leaf_node_size, size_t neighbor_size, size_t maximum_rank, T tolerance, T budget) |
void | CopyFrom (Configuration< T > &config) |
DistanceMetric | MetricType () |
size_t | ProblemSize () |
size_t | LeafNodeSize () |
size_t | NeighborSize () |
size_t | MaximumRank () |
T | Tolerance () |
T | Budget () |
bool | IsSymmetric () |
bool | UseAdaptiveRanks () |
bool | SecureAccuracy () |
Public Attributes | |
SPDMATRIX * | K = NULL |
Data< T > * | w = NULL |
Data< T > * | u = NULL |
Data< T > * | input = NULL |
Data< T > * | output = NULL |
T | lambda = 0.0 |
bool | do_ulv_factorization = true |
![]() | |
size_t | m |
size_t | max_depth |
DistData< STAR, CBLK, pair< T, size_t > > * | NN_cblk |
DistData< STAR, CIDS, pair< T, size_t > > * | NN |
vector< size_t > | morton |
SPLITTER | splitter |
Additional Inherited Members | |
![]() | |
typedef T | T |
These are data that shared by the whole local tree. Distributed setup inherits mpitree::Setup.
This class does not have to inherit DistData, but it have to inherit DistVirtualMatrix<T>ESSENTIAL: this is an abstract function MPI request Kij from rank ( j % size ) exchange ids allocate buffer for data fetch subrows exchange data ESSENTIAL: return a submatrix MPI request KIJ from rank ( j % size ) exchange ids allocate buffer for data fetch submatrix exchange data merging data overload operator
|
inline |
Shallow copy from the config.
bool hmlp::mpigofmm::Setup< SPDMATRIX, SPLITTER, T >::do_ulv_factorization = true |
whether the matrix is symmetric use ULV or Sherman-Morrison-Woodbury
Data<T>* hmlp::mpigofmm::Setup< SPDMATRIX, SPLITTER, T >::input = NULL |
buffer space, either dimension needs to be n
SPDMATRIX* hmlp::mpigofmm::Setup< SPDMATRIX, SPLITTER, T >::K = NULL |
The SPDMATRIX (accessed with gids: dense, CSC or OOC)
T hmlp::mpigofmm::Setup< SPDMATRIX, SPLITTER, T >::lambda = 0.0 |
regularization
Data<T>* hmlp::mpigofmm::Setup< SPDMATRIX, SPLITTER, T >::w = NULL |
rhs-by-n, all weights and potentials.