Skip to content
This repository has been archived by the owner on Jul 17, 2024. It is now read-only.

Latest commit

 

History

History
41 lines (29 loc) · 3.42 KB

README.md

File metadata and controls

41 lines (29 loc) · 3.42 KB
Raw

Design overview of the math library

The math library provides basic linear algerbra functionality.

Vector classes

Classes declared in Vector.h implement the basic functionality of algebraic vectors. Vectors are templated on their element type, which is typically double or float, and their orientation, which is either VectorOrientation::row or VectorOrientation::column.

There are three user-facing vector classes:

  • A ConstVectorReference does not own any memory, and instead references the memory of a Vector (see below). ConstVectorReference includes functions that do not modify the vector contents, such as Size and operator==.
  • A VectorReference still references the memory of a Vector and adds functions that modify the vector contents, like Reset and Fill.
  • A Vector owns its memory and includes all the functionality of ConstVectorReference and VectorReference.

Most vector-vector algebraic operations are implemented outside the Vector classes, and appear in VectorOperations.h.

Matrix classes

Classes declared in Matrix.h implement the basic functionality of algebraic matrices. Matrices are templated on their element type, which is typically double or float, and their layout, which is either MatrixLayout::rowMajor or MatrixLayout::columnMajor.

There are three user-facing matrix classes, which mirror the three user-facing vector classes:

  • ConstMatrixReference
  • MatrixReference
  • Matrix

Most algebraic operations involving matrices are implemented outside the Matrix classes, and appear in MatrixOperations.h.

Tensor classes

Classes declared in Tensor.h implement the basic functionality of three-dimensional algebraic matrices. In other libraries, these are sometimes called Cubes, but here we call them Tensors (although the mathematical definition of a Tensor is more broad). Tensors are templated on their element type, which is typically double or float. They also have three template parameters that control their layout, which take the values

  • Dimension::row
  • Dimension::column
  • Dimension::channel

For example, Tensor<ElementType, Dimension::channel, Dimension::column, Dimension::row> is a Tensor whose primary dimension is the channel dimension and its secondary dimension is the column dimension. As above, the three user-facing tensor classes are:

  • ConstTensorReference
  • TensorReference
  • Tensor

Most algebraic operations involving tensors are implemented outside the Tensor classes, and appear in TensorOperations.h.

Operations

As noted above, algebraic operations on vectors, matrices, and tensors appear in the VectorOperations.h, MatrixOperations.h, and TensorOperations.h files. Some of these operations have multiple implementations: a native (built-in) implementation and a BLAS implementation. Typically, the user is unaware of the underlying implementation, and uses commands like math::Multiply(s, M) (which scales the matrix M by the scalar s). If the precompiler macro USE_BLAS is defined, this command invokes the BLAS implementation and otherwise it invokes the native implementation.

To explicitly invoke a specific implementation, use math::Internal::MatrixOperations<math::ImplementationType::native>::Multiply or math::Internal::MatrixOperations<math::ImplementationType::openBlas>::Multiply. If USE_BLAS is not defined during compilation, then both of these calls will invoke the native implementation.