Skip to content

Maths library

etch mathematics and machine learning operations are supported by a header-only fully-templated C++ library.

Detailed developer documentation for the C++ implementation of the maths libraries will be available in the Fetch Ledger section in due course. Developers should be comfortable with SFINAE.

Architecture

A core component of the maths library is the tensor.hpp class which handles N-dimensional array mathematics. This is crucial for the machine learning library but can also be used for any generalised matrix algebra.

The remainder of the library contains templated free functions that can be called with the following types:

  • Built-in C++ types such as double, int, float, etc.
  • C++ tensors of built-in types such as Tensor<double>.
  • etch types such as Fixed32.
  • etch tensors types such as Tensor<Fixed32>.

The header file fundamental_operators.hpp contains common operations Add, Subtract, Multiply, and Divide.

The file matrix_operations.hpp contains Max, ArgMax, Product, Sum, Dot, etc.

The standard_functions directory contains header files for additional standard operations.

The following block diagram gives a rough indication of the library structure.

Advice

This is work in progress.

Maths library architecture

Tensor

A Tensor is a wrapper for a data_ object which is, by default, a SharedArray managed by the vectorise library. This library manages the vectorisation/SIMD on the underlying data.

Tensor objects provide interfaces for manipulating arrays at a mathematical/algebraic level while allowing implementations to be efficient and vectorisable.

Tensors have related helper classes such as TensorIterator, TensorBroadcast, and TensorSlice that permit efficient and convenient manipulation such as accessing, transposing, and slicing.

Check out the available mathematical functions in etch here.

Working with the maths library

When working with the C++ maths library, take note of the following:

  • Functions should have two interfaces: one that takes a reference to the return object, and one that creates the return object internally.

  • Function design decisions should follow Numpy conventions where possible.