Package org.apache.sysds.runtime.compress.lib

Class CLALibSort

java.lang.Object

org.apache.sysds.runtime.compress.lib.CLALibSort

public final class CLALibSort
extends Object

Method Summary

Modifier and Type	Method	Description
static MatrixBlock	sort(CompressedMatrixBlock mb, SortIndex fn)	Sort (order) a compressed matrix in place of the order built-in, while keeping the result compressed.
static MatrixBlock	sort(CompressedMatrixBlock mb, MatrixValue weights, MatrixBlock result, int k)	Compute the sorted value/weight table used by the quantile/median/IQM operations (the sort / qsort lop), exploiting compression to sort the few distinct values instead of all rows.

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

sort

public static MatrixBlock sort(CompressedMatrixBlock mb,
 SortIndex fn)

Sort (order) a compressed matrix in place of the order built-in, while keeping the result compressed. The compressed fast-path only supports the case the user can benefit from: a single column held in a single column group, sorted ascending and returning the sorted values (not the index permutation). For everything else (multiple columns, multiple column groups, descending order, index return, or a column-group encoding without a sort implementation) this returns null so the caller can fall back to a decompressed reorg.

Parameters:

mb - the compressed matrix to sort

fn - the sort specification carried by the reorg operator

Returns:

the sorted compressed matrix, or null if the compressed fast-path does not apply

sort

public static MatrixBlock sort(CompressedMatrixBlock mb,
 MatrixValue weights,
 MatrixBlock result,
 int k)

Compute the sorted value/weight table used by the quantile/median/IQM operations (the sort / qsort lop), exploiting compression to sort the few distinct values instead of all rows. The compressed fast-path applies to an unweighted sort of a single column held in a single column group. The produced table is bit-for-bit identical to MatrixBlock.sortOperations(MatrixValue, MatrixBlock, int): a (1 + nnz) x 2 matrix holding one row per non-zero value (weight 1) plus a single collapsed row for the zeros (weight = number of zeros), sorted ascending by value. For every other case (weights present, multiple columns or groups, or an encoding without a sort implementation) it falls back to a decompressed sort.

Parameters:

mb - the compressed matrix to sort

weights - optional per-row weights, or null

result - the result matrix (reused by the fallback)

k - the parallelization degree

Returns:

the sorted value/weight table