NetFabric.Numerics.Tensors 3.1.0

NetFabric.Numerics.Tensors

Dealing with SIMD in .NET for optimized code can be complex, but this library offers a practical solution. It provides a reusable and highly-optimized iterations on Span<T>, enabling the application of both pre-defined and custom operations to each element.

Using generics, the library accommodates any type embracing .NET generic math.

Within the library, you'll find pre-defined operations such as Sqrt(), Sin(), Negate(), Add(), Divide(), Multiply(), AddMultiply(), Sum(), Average(), and many more.

For custom operations, the library allows the definition of operators through interfaces like IUnaryOperator<T>, IBinaryOperator<T>, ITernaryOperator<T>, or IAggregationOperator<T>. These operators can be applied seamlessly using the Apply() or Aggregate() methods.

Documentation for this library is available at NetFabric.Numerics.Tensors Documentation.

Usage

The library includes methods tailored for operations involving one, two, or three ReadOnlySpan<T>. Results are provided in a Span<T>, with the condition that the destination Span<T> must be of the same size or larger than the sources. Inplace operations are supported when the destination parameter matches any of the sources.

For example, given a variable data of type Span<int>, the following code snippet replaces each element of the span with its square root:

TensorOperations.Sqrt(data, data);

Note that since data serves as both the source and destination, the operation is performed in-place.

Given variables x, y, and result, all of type Span<float> and of the same size, the following example updates each element in result with the sum of the corresponding elements in x and y:

TensorOperations.Add(x, y, result);

In addition to per-element operations, the library also supports various aggregation operations, including:

var sum = TensorOperations.Sum(values);
var average = TensorOperations.Average(values);
var product = TensorOperations.Product(values);
var min = TensorOperations.Min(values);
var max = TensorOperations.Max(values);
var (min, max) = TensorOperations.MinMax(values);

The library offers functionality to retrieve the index of the first element meeting a specific condition:

var indexOfEqual = IndexOfEquals(values, 0);
var indexOfGreaterThan = IndexOfGreaterThan(values, 0);
var indexOfGreaterThanOrEqual = IndexOfGreaterThanOrEqual(values, 0);
var indexOfLessThan = IndexOfLessThan(values, 0);
var indexOfLessThanOrEqual = IndexOfLessThanOrEqual(values, 0);

Please consult the documentation to get the complete list of pre-defined operations.

Custom Operations

While NetFabric.Numerics.Tensors provides various pre-defined operations, combining them might not be efficient. Custom operators can be implemented, allowing the definition of specific operations for each element of the source, while still benefiting from high-performance reusable iteration code. Please refer to the documentation for instructions on how to do it.

Benchmarks

Please consult the documentation to obtain benchmarking results.