Home

Awesome

GitHub last commit (main) Build and test Coverage NuGet Version NuGet Downloads Gitter

NetFabric.Hyperlinq

NetFabric.Hyperlinq contains alternative implementations of many operations found in the System.Linq namespace:

This implementation favors performance in detriment of assembly binary size (lots of overloads).

Contents

Fast enumeration

NetFabric.Hyperlinq can enumerate faster the results of a query than System.Linq by performing all of the following:

The performance is equivalent when the enumerator is a reference-type. This happens when the enumerable is generated using yield or when it's cast to one of the BCL enumerable interfaces (IEnumerable, IEnumerable<>, IReadOnlyCollection<>, ICollection<>, IReadOnlyList<>, IList<>, or IAsyncEnumerable<>). In the case of operation composition, this only affects the first operation. The subsequent operations will have value-type enumerators.

Reduced heap allocations

NetFabric.Hyperlinq allocates as much as possible on the stack. Enumerables and enumerators are defined as value-types. Generics constraints are used for the operation parameters so that the value-types are not boxed.

It only allocates on the heap for the following cases:

Benchmarks

The results of the benchmarks comparing multiple LINQ libraries can be found in the LinqBenchmarks repository.

The results of the benchmarks included in this repository can be found in the Benchmarks folder.

The names of the benchmarks are structured as follow:

Usage

using System;
using System.Linq;
using NetFabric.Hyperlinq; // add this directive

public static void Example(IReadOnlyList<int> list)
{
  var result = list
    .AsValueEnumerable()
    .Where(item => item > 2)
    .Select(item => item * 2);

  foreach(var value in result)
    Console.WriteLine(value);
}

This allows the use of System.Linq operators on NetFabric.Hyperlinq enumerables. OrderByDescending() is not yet available in Netfabric.Hyperlinq but can still be used without requiring any conversion:

public static void Example(IReadOnlyList<int> list)
{
  var result = list
    .AsValueEnumerable()
    .Where(item => item > 2)
    .OrderByDescending(item => item) // is not yet available in Netfabric.Hyperlinq
    .AsValueEnumerable()
    .Select(item => item * 2);

  foreach(var value in result)
    Console.WriteLine(value);
}

To add NetFabric.Hyperlinq operations after a System.Linq operation, simply add one more AsValueEnumerable() or AsAsyncValueEnumerable().

Value delegates

Calling a lambda expression for each item of the collection is very expensive. NetFabric.Hyperlinq supports an alternative that is not as practical but that has much better performance.

readonly struct MultiplyBy2
    : IFunction<int, int>
{
    public int Invoke(int element)
        => element * 2;
}

readonly struct LessThan
    : IFunction<int, bool>
{
	readonly int value;
	
    public LessThan(int value)
		=> this.value = value;
    
    public bool Invoke(int element)
        => element < value;
}

public static void Example(IReadOnlyList<int> list)
{
  var result = list
    .AsValueEnumerable()
    .Where(new LessThan(10))
    .Select<int, MultiplyBy2>();

  foreach(var value in result)
    Console.WriteLine(value);
}

The instances are allocated on the stack and the methods calls are non-virtual.

Generation operations

In NetFabric.Hyperlinq, the generation operations like Empty(), Range(), Repeat() and Return() are static methods implemented in the static class ValueEnumerable. To use them, instead of the System.Linq equivalents, simply use ValueEnumerable instead of Enumerable.

public static void Example(int count)
{
  var source = ValueEnumerable
    .Range(0, count)
    .Select(item => item * 2);

  foreach(var value in source)
    Console.WriteLine(value);
}

Composition

NetFabric.Hyperlinq operations can be composed just like with System.Linq. The difference is on how each one optimizes the internals to reduce the number of enumerators required to iterate the values.

Both System.Linq and NetFabric.Hyperlinq optimize the code in the following example so that only one enumerator is used to perform both the Where() and the Select():

var result = source.AsValueEnumerable()
    .Where(item => item > 2)
    .Select(item => item * 2);

NetFabric.Hyperlinq includes many more composition optimizations. In the following code, only one enumerator is used:

var result = array.AsValueEnumerable()
    .Skip(1)
    .Take(10)
    .Where(item => item > 2)
    .Select(item => item * 2)
    .First();

Option

In System.Linq, the aggregation operations like First(), Single() and ElementAt(), throw an exception when the source has no items. Often, empty collections are a valid scenario and exception handling is very slow. System.Linq has alternative methods like FirstOrDefault(), SingleOrDefault() and ElementAtOrDefault(), that return the default value instead of throwing. This is still an issue when the items are of a value-type, where there's no way to distinguish between an empty collection and a valid item.

In NetFabric.Hyperlinq, aggregation operations return an Option<> type. This is similar in behavior to the Nullable<> but it can contain reference types.

Here's a small example using First():

var result = source.AsValueEnumerable().First();
if (result.IsSome)
  Console.WriteLine(result.Value);

It also provides a deconstructor so, you can convert it to a tuple:

var (isSome, value) = source.AsValueEnumerable().First();
if (isSome)
  Console.WriteLine(value);

If you prefer a more functional approach, you can use Match() to specify the value returned when the collection has values and when it's empty. Here's how to use it to define the previous behavior of First() and FirstOrDefault():

var first = source.AsValueEnumerable()
  .First()
  .Match(
    item => item,
    () => throw new InvalidOperationException("Sequence contains no elements"));

var firstOrDefault = source.AsValueEnumerable()
  .First()
  .Match(
    item => item,
    () => default);

Console.WriteLine(first);
Console.WriteLine(firstOrDefault);

Match() can also be used to define actions:

source.AsValueEnumerable()
  .First()
  .Match(
    item => Console.WriteLine(item),
    () => { });

The NetFabric.Hyperlinq operations can be applied to Option<>, including Where(), Select() and SelectMany(). These return another Option<> with the predicate/selector applied to the value, if it exists.

source.AsValueEnumerable()
  .First()
  .Where(item => item > 2)
  .Match(
    item => Console.WriteLine(item),
    () => { });

Buffer pools

Buffer pools allow the use of heap memory without adding pressure to the garbage collector. It pre-allocates a chunk of memory and "rents" it as required. The garbage collector will add this memory to the Large Object Heap (LOH).

ToArray() is frequently used to cache values for a brief period and the use of buffer pools may be useful.

Netfabric.Hyperlinq adds an overload that takes a ArrayPool<> as a parameter:

void Method()
{
  using var buffer = source.AsValueEnumerable()
      .ToArray(ArrayPool<int>.Shared);
  var memory = buffer.Memory;
  // use memory here
}

It returns an instance of a IMemoryOwner<>. The using statement guarantees that it is disposed and the buffer automatically returned to the pool.

SIMD

NetFabric.Hyperlinq uses SIMD implicitly to improve performance of some operations. Many times this can only be done explicitly because it can only be used on a limited number of types and operations on them.

In NetFabric.Hyperlinq, alternative methods that use SIMD, have the word Vector at the end of its name. You'll also find that the operations that require an expression, now require two.

var result = list
  .AsValueEnumerable()
  .SelectVector(item = item * 2, item = item * 2)
  .ToArray();

SIMD improves performance by performing the same operations simultaneously on multiple items. The operation can only be performed if this number is met. This means that the remaining number of items has to be processed without SIMD. The first expression is applied on a System.Numerics.Vector<>, while the second one is applied on an item.

These methods also support the use of value delegates. In this case, the struct containing the expressions must implement two IFunction<,>:

readonly struct MultiplyBy2
    : IFunction<Vector<int>, Vector<int>>
    , IFunction<int, int>
{
    public Vector<int> Invoke(Vector<int> element)
        => element * 2;

    public int Invoke(int element)
        => element * 2;
}

public static void Example(List<int> list)
{
  var result = list
    .AsValueEnumerable()
    .SelectVector<int, MultiplyBy2>()
    .ToArray();

  foreach(var value in result)
    Console.WriteLine(value);
}

Please note that the operation SelectVector() does not return an enumerable. It returns the context required for subsequent operations like ToArray(), ToList(), and Sum().

Up until now I haven't found a way to improve the performance of Select() by using SIMD. The returned context allows the use of composition, exposing only the operations that can gain with the use of SIMD.

Documentation

Articles explaining implementation:

Supported operations

References

Credits

The following open-source projects are used to build and test this project:

License

This project is licensed under the MIT license. See the LICENSE file for more info.