Quick Links
Release Notes Getting Started Quick Reference Parallel Loops: gfor

ArrayFire is the fastest GPU matrix library with the simplest API.

ArrayFire is the fastest GPU software

It contains the fastest implementations of hundreds of matrix, signal, and image processing routines that enable it outperform CPU libraries like IPP, MKL, Eigen, Armadillo, and more.

It is optimized for any CUDA-enabled GPU. The same code will run on laptops, desktops, or servers.

It includes thousands of lines of highly-tuned device code.

It performs run-time analysis of your code to increase arithmetic intensity and memory throughput while avoiding unnecessary temporary allocations.

It combines and enhances all the best CUDA libraries available, including the fastest FFT, BLAS, and LAPACK implementations.

ArrayFire is the easiest-to-use GPU software

A simple array notation you can learn in minutes.

A few lines of ArrayFire code accomplishes what would have taken 10-100X lines in raw CUDA.

It is easier than templated programming and goes farther than simple directive-based approaches (and outperforms those approaches too).

It supports easily scaling to take advantage of multiple GPUs.

It can be used in C/C++ applications by itself or integrated with your existing CUDA code (see more).

ArrayFire is the most comprehensive GPU software

It has hundreds of functions you need to make your code faster including arithmetic, linear algebra, statistics, signal processing, image processing, and related algorithms (see more).

It supports single- and double-precision floating point values, complex numbers, booleans, 32-bit signed and unsigned integers (see more).

It supports manipulating vectors, matrices, and N-dimensional arrays (see more).

It can execute loop iterations in parallel with gfor (see more).

See it in action...

Here's a stripped down example of Monte-Carlo estimation of Pi:

    #include <stdio.h>
    #include <arrayfire.h>
    using namespace af;

    int main() {

        // create GPU arrays
        int n = 20e6; // use 20 million random samples
        array x = randu(n), y = randu(n);

        // determine how many fell inside the unit circle?
        // ... calculate on GPU, return result to CPU
        float pi = 4.0 * sum<float>(hypot(x,y) < 1) / n;

        printf("pi = %g\n", pi);
        return 0;
    }

Download and Requirements

Download the latest stable version of the library and view documentation online. See Release Notes for a list of changes in each version. You can also download nightly builds including the latest features and bug fixes.

Supported Platforms:

Windows (32 and 64-bit) - XP, Vista, 7
Linux (32 and 64-bit) - Ubuntu 10+, Fedora 10+, OpenSUSE 11+, RHEL 5+, CentOS 5+, SLES 10+
Mac OSX (64-bit) - Snow Leopard (10.6.x)

Requirements:

CUDA-enabled GPU
CUDA 4.1 toolkit and driver
Windows requires Visual Studio 2008 or 2010 and CUDA 4.1.28
Linux and Mac require gcc/g++ versions 4.2 through 4.5
- gcc 4.6 requires a workaround

The Getting Started tutorial walks through more detailed steps on getting your first example to compile and run. You can also browse through the examples online.

Support