Intel® oneAPI Data Analytics Library Developer Guide and Reference

Developer Guide and Reference

  • 2021.4
  • 09/27/2021
  • Public Content
Contents

Essential Interfaces for Algorithms

In addition to Generic Interfaces, more methods enable interfacing numeric tables with algorithms.
The getDataLayout method provides information about the data layout:
Data Layout
Description
soa
Structure-Of-Arrays (SOA). Values of individual data features are stored in contiguous memory blocks.
aos
Array-Of-Structures (AOS). Feature vectors are stored in contiguous memory block.
csr_Array
Condensed-Sparse-Row (CSR).
lowerPackedSymetricMatrix
Lower packed symmetric matrix
lowerPackedTriangularMatrix
Lower packed triangular matrix
upperPackedSymetricMatrix
Upper packed symmetric matrix
upperPackedTriangularMatrix
Upper packed triangular matrix
unknown
No information about data layout or unsupported layout.
Rather than access the entire in-memory data set, it is often more efficient to process it by blocks. The key methods that oneDAL algorithms use for per-block data access are
getBlockOfRows()
 and
getBlockOfColumnValues()
. The
getBlockOfRows()
 method accesses a block of feature vectors, while the
getBlockOfColumnValues()
 method accesses a block of values for a given feature. A particular algorithm uses
getBlockOfRows()
,
getBlockOfColumnValues()
, or both methods to access the data. The efficiency of data access highly depends on the data layout and on whether the data type of the feature is natively supported by the algorithm without type conversions. Refer to the Performance Considerations section in the description of a particular algorithm for a discussion of the optimal data layout and natively supported data types.
When the data layout fits the per-block data access pattern and the algorithm requests the data type that corresponds to the actual data type, the
getBlockOfRows()
 and
getBlockOfColumnValues()
 methods avoid data copying and type conversion. However, when the layout does not fit the data access pattern or when type conversion is required, both methods automatically re-pack and convert data as required.
When dealing with custom or unsupported data layouts, you must implement NumericTableIface, DenseNumericTableIface interfaces, and optionally CSRNumericTableIface or PackedNumericTableIface interfaces.
Some algorithms, such as Moments of Low Order, compute basic statistics (minimums, maximums, and so on). The other algorithms, such as Correlation and Variance-Covariance Matrices or Principal Component Analysis, require some basic statistics on input. To avoid duplicated computation of basic statistics, oneDAL provides methods to store and retrieve basic statistics associated with a given numeric table:
basicStatistics.set()
 and
basicStatistics.get()
. The following basic statistics are computed for each numeric table:
  • minimum - minimum
  • maximum - maximum
  • sum - sum
  • sumSquares - sum of squares
The default data type of basic statistics is float.
Special Interfaces for the HomogenNumericTable and Matrix Classes
  • Use the assign method to initialize elements of a dense homogeneous numeric table with a certain value, that is, to set all elements of the matrix to zero.
  • Use the operator [] method to access rows of a homogeneous dense numeric table.
Special Interfaces for the PackedTriangularMatrix and PackedSymmetricMatrix Classes
  • While you can use generic
    getArray()
     and
    setArray()
     methods to access the data in a packed format, in algorithms that have specific implementations for a packed data layout, you can use more specific
    getPackedValues()
     and
    releasePackedValues()
     methods.
Special Interfaces for the CSRNumericTable Class
  • To access three CSR arrays (values , columns, and rowIndex), use
    getArrays()
     and
    setArrays()
     methods instead of generic
    getArray()
     and
    setArray()
     methods. For details of the arrays, see CSR data layout.
  • Similarly, in algorithms that have specific implementations for the CSR data layout, you can use more specific
    getBlockOfCSRValues()
     and
    releaseBlockOfCSRValues()
     methods.
Special Interfaces for the MergedNumericTable Class
  • To add a new array to the object of the MergedNumericTable class, use the
    addNumericTable()
     method.

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.