ai.rapids.cudf

Class Aggregation128Utils

java.lang.Object

ai.rapids.cudf.Aggregation128Utils

public class Aggregation128Utils
extends Object

Utility methods for breaking apart and reassembling 128-bit values during aggregations to enable hash-based aggregations and detect overflows.

Constructor Summary

Constructors

Constructor and Description
Aggregation128Utils()

Method Summary

Modifier and Type	Method and Description
static Table	combineInt64SumChunks(Table chunks, DType type) Reassemble a column of 128-bit values from a table of four 64-bit integer columns and check for overflow.
static ColumnVector	extractInt32Chunk(ColumnView col, DType outType, int chunkIdx) Extract a 32-bit chunk from a 128-bit value.

Methods inherited from class java.lang.Object

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

Constructor Detail

Aggregation128Utils

public Aggregation128Utils()

Method Detail

extractInt32Chunk

public static ColumnVector extractInt32Chunk(ColumnView col,
                                             DType outType,
                                             int chunkIdx)

Extract a 32-bit chunk from a 128-bit value.

Parameters:

col - column of 128-bit values (e.g.: DECIMAL128)

outType - integer type to use for the output column (e.g.: UINT32 or INT32)

chunkIdx - index of the 32-bit chunk to extract where 0 is the least significant chunk and 3 is the most significant chunk

Returns:

column containing the specified 32-bit chunk of the input column values. A null input row will result in a corresponding null output row.

combineInt64SumChunks

public static Table combineInt64SumChunks(Table chunks,
                                          DType type)

Reassemble a column of 128-bit values from a table of four 64-bit integer columns and check for overflow. The 128-bit value is reconstructed by overlapping the 64-bit values by 32-bits. The least significant 32-bits of the least significant 64-bit value are used directly as the least significant 32-bits of the final 128-bit value, and the remaining 32-bits are added to the next most significant 64-bit value. The lower 32-bits of that sum become the next most significant 32-bits in the final 128-bit value, and the remaining 32-bits are added to the next most significant 64-bit input value, and so on.

Parameters:

chunks - table of four 64-bit integer columns with the columns ordered from least significant to most significant. The last column must be of type INT64.

type - the type to use for the resulting 128-bit value column

Returns:

table containing a boolean column and a 128-bit value column of the requested type. The boolean value will be true if an overflow was detected for that row's value when it was reassembled. A null input row will result in a corresponding null output row.