Get Started

Neanderthal’s default option is to use use native libraries, so it is very important that you do not skip any part of this guide.

How to Get Started

Installation (Leiningen)

The most straightforward way to include Neanderthal in your project is with Leiningen. Check the Hello World project out for the complete example. Please note that, if you need an older version of Neanderthal, they may need a bit more specific installation steps, which are explained in Requirements.

  • Add the following dependency to your project.clj,:
  • Add a MKL distribution jar [org.bytedeco/mkl "2024.0-1.5.10" :classifier linux-x86_64-redist] as your project’s dependency.
  • (optional) Add a CUDA distribution jar [org.bytedeco/cuda "12.3-8.9-1.5.10" :classifier linux-x86_64-redist] as your project’s dependency.

Neanderhtal will use the native CPU MKL (and/or CUDA) binaries from that jar automatically, so you don’t need to do anything else. If the jar is not present, Neanderthal will expect you to have a system-wide MKL installation as explained in Native Engine Requirements.33 Note: MKL distribution size is 200 MB, while CUDA is 2.0 GB! Lein will download these JARs the first time you include them, which might take some time, so it’s a good idea to run lein deps and wait each time you update the version.

Also note the :classifier, which is OS dependent. You may omit it, but then binaries for all operating systems will be downloaded, which will take longer. For, windows, use windows-x86_64-redist, and macosx-x86_64-redist for mac. Sadly, macOS is currently only supported on Intel CPUs, and you’ll have to use LEGACY Neanderthal versions, 0.47.0 and older, so it can be a bit messy to set up.

Usage

First use or require uncomplicate.neanderthal.core and uncomplicate.neanderthal.native in your namespace, and you’ll be able to call appropriate functions from the Neanderthal library.

(ns example
  (:use [uncomplicate.neanderthal core native]))

Now you can create neanderthal vectors and matrices, and do computations with them.

Here is how we create two double-precision vectors and compute their dot product:

(def x (dv 1 2 3))
(def y (dv 10 20 30))
(dot x y)

This is one of the ways to multiply matrices:

(def a (dge 3 2 [1 2 3 4 5 6]))
(def b (dge 2 3 [10 20 30 40 50 60]))
(mm a b)

Overview and Features

Neanderthal is a Clojure library for fast matrix and linear algebra computations that supports pluggable engines:

  • The native engine is based on a highly optimized native Intel’s MKL library of BLAS and LAPACK computation routines (MKL is not open-source, but it is free to use and redistribute since 2016).
  • The CUDA GPU engine is based on cuBLAS and supports all modern Nvidia GPUs. It uses ClojureCUDA and JCuda libraries. Check out Uncomplicate ClojureCUDA.
  • The OpenCL GPU engine is based on OpenCL BLAS routines from CLBlast library for even more computational power when needed. It uses ClojureCL and JOCL libraries. Check out Uncomplicate ClojureCL.

Implemented Features

  • Data structures: double and single precision vectors, dense matrices (GE), triangular matrices (TR), symmetric matrices (SY), banded, diagonal, etc.;
  • BLAS Level 1, 2, and 3 routines;
  • Lots of (but not all) LAPACK routines;
  • Sparse matrices and vectors.
  • Various Clojure vector and matrix functions (transpositions, submatrices etc.);
  • Easy and efficient data mapping and transfer to and from GPUs;
  • Fast map, reduce and fold implementations for the provided structures.
  • OpenCL GPU support
  • CUDA GPU support

On the TODO List

  • ~”Tensors”~: This is already available in Deep Diamond!
  • Support for complex numbers;

Requirements

You need at least Java 8, but the newer the better.

LEGACY (0.47.0 and older): Open JVM modules.

If you are running on Java 9 or higher, you need to enable the java.base module. Add the following to your JVM options (:jvm-opts in leiningen): "--add-opens=java.base/jdk.internal.ref=ALL-UNNAMED" and "--add-opens=java.base/sun.nio.ch=ALL-UNNAMED".

Neanderthal’s data structures are written in Clojure, so many functions work even without native engines. However, you probably need Neanderthal because of its fast BLAS native or GPU engines. Here is how to make sure they are available.

GPU drivers for the OpenCL GPU engine

Everything will magically work (no need to compile anything) on Nvidia, AMD, and Intel’s GPUs and CPUs as long as you have appropriate GPU drivers.

Works on Linux, Windows, and OS X!

Follow the ClojureCL getting started guide for the links for the GPU platform that you use and more detailed info.

If you use a pre-2.0 OpenCL platform (Nvidia and/or OS X), you’ll have to use command-queue-1 and/or with-default-1 instead of command-queue and with-default.

GPU drivers for the CUDA GPU engine

Everything will magically work (no need to compile anything) on Nvidia, provided that you have Nvidia drivers, and included the appropriate Nvidia Toolkit JavaCPP jar ([org.bytedeco/cuda "12.3-8.9-1.5.10" :classifier linux-x86_64-redist]).

Follow the ClojureCUDA getting started guide for the links for the GPU platform that you use and more detailed info.

macOS doesn’t support CUDA 11 and higher (and Apple hasn’t shipped Nvidia GPUs since 2014 anyway). You’ll have to exclude CUDA dependency in your build script (see the the Hello World project).

If you’re using a LEGACY version (0.47.0 and older) you’ll need to install the CUDA Toolkit on your machine, instead of using JavaCPP CUDA binaries.

LEGACY (0.47.0 and older) The native library used by Neanderthal’s native engine (Optional)

The following is not needed if you include [org.bytedeco/mkl-platform-redist “2020.3-1.5.4”] dependency. (Also, ignore this section completely if you’re using 0.48.0 and up.)

This section deals with system-wide MKL installation.

  • Works on Linux, OS X, and Windows!

Neanderthal uses the native Intel MKL library and expects that you make it available on your system, typically as shared xyz.so, xyz.dll, or xyz.dylib files. Intel MKL is highly optimized for various architectures; its installation comes with many optimized binaries for all supported architectures, that are then selected during runtime according to the hardware at hand.

You do not need to compile or tune anything yourself.

These are alternative ways to make MKL available on your system globally; either:

  1. Install it through a package manager it that is available on your system. I use Arch Linux (pacman -Syu intel-mkl), and I believe that many other Linux distributions now ship MKL in their repositories..

OR:

  1. (Optionally) Install MKL using a GUI installer provided by Intel free of charge. In case you use this method, you may set environment variables as explained in this guide, but it is probably not required, since you do not need to compile anything.
  2. Put all required binary files (that you installed with MKL installer or copied from wherever you acquired them) in a directory that is available from your LD_LIBRARY_PATH (Linux), DYLD_LIBRARY_PATH (OSX) or, PATH (windows). Those binary files are available from anyone who installed MKL and have the (free) license to redistribute it with a project.

This is the list of MKL files that Neanderthal requires:

  • Mandatory MKL libraries (this is what I use on my Intel Core i7 4790k):
    • libmkl_rt.so
    • libmkl_core.so
    • libmkl_intel_lp64.so
    • libiomp5.so
    • libmkl_intel_thread
    • libmkl_avx2.so (if your CPU supports AVX2, which it probably does)
  • Optionally, your processor might support additional set of instructions that may require an additional file from MKL. See the MKL guide and browse all available files once you install MKL to discover this. For example, I guess that you might also need libmkl_avx512.so if you have a Xeon processor.

Please note that, if you use Windows or OS X, the binary file extensions are not .so, but .dll and dylib respectively.

Note for OSX users: MKL installation on my OSX 10.11 placed libiomp5.dylib in a different folder than the rest of the MKL libraries. In such case, copy that file where it is visible, or don’t rely on the MKL installation, but select the needed library files and put them somewhere on the DYLD_LIBRARY_PATH. In newer versions of OSX, you’d have to configure the “system integrity protection (SIP)” settings for DYLD_LIBRARY_PATH to be respected by the system see more here. If you want a quick & dirty solution without much fuss, copying the dylib files and pasting them into /usr/local/lib has been reported to work by multiple users.

Note for Windows users: MKL installation on my Windows 10 keeps all required .dll files in the <install dir>\redist folder. The usual folders that keep .so and dylib on Linux and OSX, keep .lib files on Windows - you do not need those. Add the folder that contains the dlls into the PATH environment vairable, and you’re ready to go. Some Windows users reported that libiomp5.dll too is in another folder; see the note for OSX users and take the equivalent Windows action.

Final note If you prefer zero-install, just include [org.bytedeco/mkl-platform-redist "2020.3-1.5.4"] as a dependencly in your leiningen project and none of these is necessary.

Where to Go Next

Hopefully this guide got you started and now you’d like to learn more. I expect to build a comprehensive base of articles and references for exploring this topic, so please check the All Guides page from time to time. Of course, you should also check the Neanderthal API for specific details, and feel free to take a glance at the source while you are there.

It is also a good idea to follow my blog at dragan.rocks since I’ll write about Neanderthal there.

Tell Us What You Think!

Please take some time to tell us about your experience with the library and this site. Let us know what we should be explaining or is not clear enough. If you are willing to contribute improvements, even better!