Package org.apache.sysds.runtime.data
package org.apache.sysds.runtime.data
-
ClassDescriptionThis DenseBlock is an abstraction for different dense, row-major matrix formats.Dense Large Row Blocks have multiple 1D arrays (blocks), which contain complete rows.This SparseBlock is an abstraction for different sparse matrix formats.SparseBlock implementation that realizes a traditional 'coordinate matrix' representation, where the entire sparse block is stored as triples in three arrays: row indexes, column indexes, and values, where row indexes and colunm indexes are sorted in order to allow binary search.SparseBlock implementation that realizes a traditional 'compressed sparse column' representation, where the entire sparse block is stored as three arrays: ptr of length clen+1 to store offsets per column, and indexes/values of length nnz to store row indexes and values of non-zero entries.SparseBlock implementation that realizes a traditional 'compressed sparse row' representation, where the entire sparse block is stored as three arrays: ptr of length rlen+1 to store offsets per row, and indexes/values of length nnz to store column indexes and values of non-zero entries.SparseBlock implementation that realizes a 'modified compressed sparse column' representation, where each compressed column is stored as a separate SparseRow object which provides flexibility for unsorted column appends without the need for global reshifting of values/indexes but it incurs additional memory overhead per column for object/array headers per column which also slows down memory-bound operations due to higher memory bandwidth requirements.SparseBlock implementation that realizes a 'modified compressed sparse row' representation, where each compressed row is stored as a separate SparseRow object which provides flexibility for unsorted row appends without the need for global reshifting of values/indexes but it incurs additional memory overhead per row for object/array headers per row which also slows down memory-bound operations due to higher memory bandwidth requirements.Base class for sparse row implementations such as sparse row vectors and sparse scalars (single value per row).A sparse row vector that is able to grow dynamically as values are appended to it.A
TensorBlockis the most top level representation of a tensor.This represent the indexes to the blocks of the tensor.