Math.NET Numerics | Math.NET Project | GitHub

Intel Math Kernel Library (MKL)

Math.NET Numerics is designed such that performance-sensitive algorithms can be swapped with alternative implementations by the concept of providers. There is currently only a provider for linear algebra related routines, but there are plans to add additional more e.g. related to nonlinear optimization problems or signal processing.

Providers become interesting when they can leverage a platform-native high performance library like Intel MKL instead of the default purely managed provider. Math.NET Numerics provides such a provider as NuGet packages:

Since these native libraries can become very big, there are also variants supporting only a single platform, for example:

In order to leverage the MKL linear algebra provider, we need to make sure the .NET runtime can find the native libraries (see below) and then enable it by calling:


Alternatively you can also enable it by setting the environment variable MathNetNumericsLAProvider=MKL.

You can also explicitly disable the MKL provider by forcing it to use the managed provider by calling:


You can tell what provider is effectively loaded by calling Control.LinearAlgebraProvider.ToString(), which will return something along the lines of Intel MKL (x86; revision 7).

Native Binaries

In .Net, the fusion engine is responsible for finding referenced assemblies in the file system and loading them into the executing process. However, native binaries like our MKL provider are platform specific, so we need to load them with services of the platform instead of the .Net runtime. We use P/Invoke to talk to the binaries, but for this to work they must have already been loaded or the platform service needs to be able to find and load them on its own.

In order to make providers easier to use, since v3.6.0 Math.NET Numerics first tries to load native providers from a set of known directories before falling back to the platform's default behavior. In each of these directories it first looks for a processor-architecture specific folder within the directory, before looking at the directory itself:

  1. If Control.NativeProviderPath is set: {NativeProviderPath}/{Platform}/
  2. If Control.NativeProviderPath is set: {NativeProviderPath}/
  3. {AppDomain.BaseDirectory}/{Platform}/
  4. {AppDomain.BaseDirectory}/
  5. {ExecutingAssemblyPath}/{Platform}/
  6. {ExecutingAssemblyPath}/
  7. Fall back to the platform's default behavior (see below)

Where {Platform} can be one of the following: x86, x64, ia64, arm or arm64.

This means that you can, for example, place the 32 bit MKL provider binaries into C:\MKL\x86 and the 64 bit ones into C:\MKL\x64, and then set Control.NativeProviderPath = @"C:\MKL";. Numerics will automatically choose the right one depending on whether your process is running in 32 or 64 bit mode, and there is no more need to copy the large binaries to the output folder of every script or project.

Default Behavior on Windows

On Windows it is usually enough to make sure the native libraries are in the same folder as the executable. Reference the appropriate NuGet package and set "Copy to Output Directory" for both libMathNetNumericsMKL.dll and libiomp5md.dll to "Copy always", or place the two native DLLs manually into the same directory as your application's executable. There is no need to set the native provider path explicitly.

For more details how the platform default behavior works and what influences it, see Dynamic-Link Library Search Order.

Default Behavior on Linux

Native assembly resolving is very different on Linux than on Windows, simply putting the native libraries into the same folder as the executable is not enough. The safe way is to edit /etc/ and use ldconfig to tell where to look for the libraries. Alternatively you could add the path to LD_LIBRARY_PATH or even just copy them to /usr/lib.

For details see Mono's Interop with Native Libraries.

Default Behavior on Mac OS X

You can configure the search path on one of the environment variables like DYLD_LIBRARY_PATH or just copy them e.g. to /usr/lib.

For details see Mono's Interop with Native Libraries.

To build the MKL native provider for OSX:

  1. Make sure you've a valid Intel MKL licence installed on your mac (look at opt/intel). If not, you can get a free trial on intel's web site.
  2. Open the terminal
  3. cd to the folder mathnet-numerics/src/NativeProviders/OSX
  4. Run the .sh script by typing sh
  5. ... wait for the build
lionel:~ Lionel$ cd /Users/Lionel/Public/Git/GitHub/mathnet-numerics/src/NativeProviders/OSX
lionel:OSX Lionel$ ls
lionel:OSX Lionel$ sh

Check the /x86 and /x64 folders in mathnet-numerics/out/MKL: you should now find the libiomp5.dylib and libMathNetNumericsMKL.dll libaries. You need to add the path to the generated libraries in your DYLD_LIBRARY_PATH environment variable (which you can move to the folder of you choice before). To do that, open your /Users/Lionel/ file with a text editor and add the following statements.

export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:/Users/Lionel/../mathnet-numerics/out/MKL/OSX/x64
export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:/Users/Lionel/../mathnet-numerics/out/MKL/OSX/x86

Of course replace Lionel by your account login.

Have a look a the example down this page to compare MKL-provider vs. managed-provider performances.

F# Interactive

In F# Interactive, the easiest way to use native providers is to copy them to a shared directory somewhere and use them directly from there:

Control.NativeProviderPath <- @"C:\MKL"

If you are using the F# Power Tools in VisualStudio, you can also let it generate "Reference scripts for F# Interactive" right from the context menu. This will generate a script called load-references.fsx in a Scripts folder, which you can extend as follows to load the MKL provider automatically.

open System.IO
open MathNet.Numerics

Control.NativeProviderPath <- Path.Combine(__SOURCE_DIRECTORY__,"../")

This script assumes that the MKL binaries have been copied to the project directory, which is also where the NuGet packages place them by default. If you place them somewhere else, adapt the path accordingly.

See also Loading Native DLLs in F# Interactive for more alternatives.

LINQPad and assembly shadowing

The automatic strategy may still work if assembly shadowing is involved, but it often simpler and more reliable to provide the folder explicitly. This also works well in LINQPad, with and without assembly shadowing:

Control.NativeProviderPath = @"C:\MKL";

Example: Intel MKL on Linux with Mono

We also provide MKL NuGet package for Linux if you do not want to build them yourself. Assuming you have Mono and NuGet installed (here v3.2.8), you can fetch the MKL package of the right architecture (x64 or x86, uname -m if you don't know) as usual:

mono nuget.exe install MathNet.Numerics -Pre -OutputDirectory packages
mono nuget.exe install MathNet.Numerics.MKL.Linux-x64 -Pre -OutputDirectory packages

Native assembly resolving is very different on Linux than on Windows, simply putting the native libraries into the same folder as the executable is not enough. The safe way is to edit /etc/ and use ldconfig to tell where to look for the libraries, but for now we'll just copy them to /usr/lib:

sudo cp packages/MathNet.Numerics.MKL.Linux-x64.1.3.0/content/ /usr/lib/
sudo cp packages/MathNet.Numerics.MKL.Linux-x64.1.3.0/content/libMathNetNumericsMKL.dll /usr/lib/

Then we're all set and can just call Control.UseNativeMKL() if we want to use the native provider. Let's create the following C# file Example.cs:

using System;
using System.Diagnostics;
using MathNet.Numerics;
using MathNet.Numerics.LinearAlgebra;

class Program
    static void Main(string[] args)
        // Using managed code only

        var m = Matrix<double>.Build.Random(500, 500);
        var v = Vector<double>.Build.Random(500);

        var w = Stopwatch.StartNew();
        var y1 = m.Solve(v);

        // Using the Intel MKL native provider

        var y2 = m.Solve(v);

Compile and run:

# single line:
mcs -optimize -lib:packages/MathNet.Numerics.3.0.0-alpha8/lib/net40/
              -r:MathNet.Numerics.dll Example.cs -out:Example
# launch:
mono Example

Licensing Restrictions

The MKL Provider is available for free under the terms of the Intel Simplified Software License.

namespace Microsoft.FSharp.Control
namespace System
namespace System.IO
type Path = static member ChangeExtension : path: string * extension: string -> string static member Combine : path1: string * path2: string -> string + 3 overloads static member EndsInDirectorySeparator : path: ReadOnlySpan<char> -> bool + 1 overload static member GetDirectoryName : path: ReadOnlySpan<char> -> ReadOnlySpan<char> + 1 overload static member GetExtension : path: ReadOnlySpan<char> -> ReadOnlySpan<char> + 1 overload static member GetFileName : path: ReadOnlySpan<char> -> ReadOnlySpan<char> + 1 overload static member GetFileNameWithoutExtension : path: ReadOnlySpan<char> -> ReadOnlySpan<char> + 1 overload static member GetFullPath : path: string -> string + 1 overload static member GetInvalidFileNameChars : unit -> char [] static member GetInvalidPathChars : unit -> char [] ...
<summary>Performs operations on <see cref="T:System.String" /> instances that contain file or directory path information. These operations are performed in a cross-platform manner.</summary>
Path.Combine([<System.ParamArray>] paths: string []) : string
Path.Combine(path1: string, path2: string) : string
Path.Combine(path1: string, path2: string, path3: string) : string
Path.Combine(path1: string, path2: string, path3: string, path4: string) : string