Developer Guide and Reference

  • 2021.1
  • 12/04/2020
  • Public Content
Contents

Types of Numeric Tables

SYCL* Numeric Tables

SyclNumericTable
class is designed to allow user to hold processed data on the device side and avoid extra data transfer between host and device. If one tries to use traditional (i.e. no
SYCL*
in the name) numeric tables with GPU algorithm, the data transfer will occur every time the algorithm needs to get access to the data.
Generally, a SYCL* numeric table is a wrapper around regular SYCL* buffer acting as an adapter. It enables the user to call GPU algorithms without unnecessary data transfer.
SYCL* Homogeneous Numeric Table
For now, only
SyclHomogenNumericTable
is implemented for DPC++ interfaces. It has similar data layout as a traditional
HomogenNumericTable
and can be initialized, operated on and uninilialized like a traditional one. Additional capabilities of
SyclHomogenNumericTable
are described in the sections below.
Initialize
A SYCL* Homogeneous numeric table can be constructed in two ways:
  • as a traditional
    HomogenNumericTable
    from host’s CPU memory (see Homogeneous Numeric Tables),
  • by a one-dimentional
    cl::sycl::buffer
    , in which case, the numeric table will hold a reference to the obtained SYCL* buffer.
cl::sycl::buffer<float, 1> bf { data, cl::sycl::range<1>{ rows*cols } }; // data is float* array with rows*cols elements // some operations... auto tablePtr = SyclHomogenNumericTable::create(bf, cols, rows);
Operate
You can get underlying reference to SYCL* buffer from
SyclHomogenNumericTable
using
getBlockOfRows()
and
getBlockOfColumns()
methods.
auto blockDescriptor = tablePtr->getBlockOfRows(0, rows, readWrite); // blockDescriptor is an object of BlockDescriptor<float> class auto readBuffer = blockDescriptor.getBuffer().toSycl(); // readBuffer is cl::sycl::buffer<float, 1> // some operations with readBuffer ... tablePtr->releaseBlockOfRows(blockDescriptor);
SyclHomogenNumericTable
does not own
cl::sycl::buffer
it was created from. Any changes of data into this buffer will affect the related numeric table. However, the buffer that you get from numeric table using
getBlockOfRows()
or
getBlockOfColumns()
methods may be another one. In this case, you should synchronize changes of data in buffer you got to numeric table by
releaseBlockOfRows()
method call.

Heterogeneous Numeric Tables

Heterogeneous numeric tables enable you to deal with data structures that are of different data types by nature. oneDAL provides two ways to represent non-homogeneous numeric tables: AOS and SOA.
AOS Numeric Table
AOS Numeric Table provides access to observations (feature vectors) that are laid out in a contiguous memory block:
Examples
SOA Numeric Table
SOA Numeric Table provides access to data sets where observations for each feature are laid out contiguously in memory:
Examples

Homogeneous Numeric Tables

Use homogeneous numeric tables, that is, objects of the
HomogenNumericTable
class, and matrices, that is, objects of the
Matrix
,
PackedTriangularMatrix
, and
PackedSymmetricMatrix
classes, when all the features are of the same basic data type. Values of the features are laid out in memory as one contiguous block in the row-major order, that is,
Observation 1
,
Observation 2
, and so on. In oneDAL,
Matrix
is a homogeneous numeric table most suitable for matrix algebra operations.
For triangular and symmetric matrices with reduced memory footprint, special classes are available:
PackedTriangularMatrix
and
PackedSymmetricMatrix
. Use the DataLayout enumeration to choose between representations of triangular and symmetric matrices:
  • Lower packed:
    lowerPackedSymetricMatrix
    or
    lowerPackedTriangularMatrix
  • Upper packed:
    upperPackedTriangularMatrix
    or
    upperPackedSymetricMatrix

CSR Numeric Table

oneDAL offers the
CSRNumericTable
class for a special version of a homogeneous numeric table that encodes sparse data, that is, the data with a significant number of zero elements. The library uses the Condensed Sparse Row (CSR) format for encoding:
Three arrays describe the sparse matrix M as follows:
  • The array values contains non-zero elements of the matrix row-by-row.
  • The j-th element of the array columns encodes the column index in the matrix M for the j-th element of the array values.
  • The i-th element of the array rowIndex encodes the index in the array values corresponding to the first non-zero element in rows indexed i or greater. The last element in the array rowIndex encodes the number of non-zero elements in the matrix M.
The library supports 1-based CSR encoding only. In C++ you can specify it by providingoneBased value through the indexing parameter of type
CSRIndexing
in the constructor of
CSRNumericTable
.
Examples

Merged Numeric Table

oneDAL offers the
MergedNumericTable
class for tables that provides access to data sets comprising several logical components, such as a set of feature vectors and corresponding labels. This type of tables enables you to read those data components from one data source. This special type of numeric tables can hold several numeric tables of any type but
CSRNumericTable
. In a merged numeric table, arrays are joined by columns and therefore can have different numbers of columns. In the case of different numbers of rows in input matrices, the number of rows in a merged table equals LaTex Math image. , where LaTex Math image. is the number of rows in the i-th matrix, LaTex Math image. .
Examples

Product and Performance Information

1

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