Julia wrapper for CUDA runtime API



Build status (Linux x86-64):

Julia 0.5 Julia 0.6

Code coverage: Coverage Status

This package wraps the CUDA runtime API. For a wrapper of the lower-level driver API, see CUDAdrv.

Platform support

This has been tested on Linux, OSX, and Windows. With Windows, at least Visual Studio 2010/2012/2013/2015 are supported.


First, you need to have an NVIDIA GPU device in your computer (one that is available for computation, i.e., most likely not your graphics card), and the CUDA library installed. You have to perform these steps manually. Choose either 32-bit or 64-bit versions to match your Julia installation.

Install the Julia package using:


During installation, it should compile a couple of files in the deps/ directory. These files provide utility functions necessary for certain functionality in this package. If the build step fails, try fixing the problems and running Pkg.build("CUDArt") manually.

After installation, it's probably a good idea to run the test/runtests.jl script to find out whether everything is working on your system, or just say Pkg.test("CUDArt").

In case of errors, one thing to check is your CUDA installation itself. For example, examine whether the *.ptx files are present in deps/ and test/; look at those files and make sure they seem appropriate. (E.g., if your computer is 64-bit, are they compiled for 64-bit?)


Start by saying using CUDArt, or import CUDArt if you prefer to qualify everything with the module name. For most use cases, you'll also need to install and import the CUDAdrv package, which among other things provides functionality to launch kernels.

GPU initialization

One or more GPUs can be initialized, used for computations, and freed for other uses. There are some complexities in this process due to the interaction with Julia's garbage collection---a CUDA array object allocated in one "session" should not be usable if you close the device and then open a new "session." Fortunately, CUDArt should make the process transparent, and as a user you shouldn't have to think about this at all.

The easiest way to ensure that you get full functionality, with proper cleanup of resources, is by using the do block syntax:

result = devices(dev->true) do devlist
    # Code that does GPU computations

The argument to devices is a function that accepts an integer input (the integer representing the CUDA device, starting with 1) and returns true or false, indicating whether the device should or should not be used, respectively. dev->true means that very device will be used. The devlist variable will be defined inside the block, and is a Vector{Int} of the available devices.

If you need to make sure that only devices with sufficient capabilities are used, then use a construct like this:

result = devices(dev->capability(dev)[1]>=2) do devlist
    # Code that does GPU computations

This will select all devices that have a major capability of 2 or higher. You can query any of the properties of your device; see the device_properties and attribute functions and the list of fields. If you want to restrict your computations to just one device (perhaps leaving other devices for other users), use the nmax keyword:

result = devices(func, nmax=1) do devlist
    # Code that does GPU computations

Finally, you can request only those devices that are not busy with other tasks using:

result = devices(func, status=:free) do devlist
    # Code that does GPU computations

You can wait for specific devices to become available with wait_free(devlist).

The do block syntax initializes the devices and loads some utility functions (defined in deps/utils.cu) onto each GPU; it also ensures proper freeing of memory and unloading of code when the do block finishes. Should you want to initialize the utilities manually, you can do so by calling CUDArt.init(devlist) and CUDArt.close(devlist) where devlist is an integer device number or a list of them, e.g. 0 or [0,1]. This can be handy in case of trouble, because unfortunately the do syntax does not usually result in ideal backtraces.

If your work doesn't require any of the utility functions, you can manually manage the device:

# Code that does GPU computations

where dev is the integer device number.

Choosing/querying the active device

At any point in your code, the command device(dev) makes dev the active device. For example, commands that allocate device memory will be executed on whichever device is currently active.

Calling dev = device() will return the currently-active device


Device arrays

CUDArt supports two main types of device arrays: CudaArrays and CudaPitchedArrays. These correspond to contiguous memory blocks and "pitched pointers", respectively.

To declare an uninitialized array on the device, use:

d_A = CudaArray(Float64, (200,300))
d_B = CudaPitchedArray(Int32, (15, 40, 27))

The d_ is a conventional way of reminding yourself that the array is allocated on the device. To copy a host array to the device, use any of

d_A = CudaArray(A)
d_AP = CudaPitchedArray(A)
copy!(d_A, A)
copy!(d_AP, A)

To copy a device array back to the host, use either of

A = to_host(d_A)
copy!(A, d_A)

Most of the typical Julia functions, like size, ndims, reinterpret, eltype, fill!, etc., work on CUDA array types. One noteworthy omission is that you can't directly index a CUDA array: d_A[2,4] will fail. This is not supported because host/device memory transfers are relatively slow, and you don't want to write code that (on the host side) makes use of individual elements in a device array. If you want to inspect the values in a device array, first use to_host to copy it to host memory.

You can find out which device is storing an array using:

dev = device(d_A)

Host arrays

Another important array type is the HostArray, which is allocated by the CUDA library using pinned memory:

h_A = HostArray(Float32, (1000,1200))

There are circumstances where using a HostArray may improve the speed of memory transfers, or allow asynchronous operations using Streams.

Warning: using a HostArray in conjunction with a large memory-mapped file has been observed to cause segfaults; at the present time there is no known workaround.

Modules and custom kernels

This will not teach you about CUDA programming; for that, please refer to the CUDA documentation and other online sources. You can find an example file in deps/utils.cu.

Compiling your own modules

You can write and use your own custom kernels, first writing a .cu file and compiling it as a ptx module. On Linux, compilation would look something like this:

nvcc -ptx mycudamodule.cu

You can specify that the code should be compiled for compute capability 2.0 devices or higher using:

nvcc -ptx -gencode=arch=compute_20,code=sm_20 mycudamodule.cu

If you want to write code that will support multiple datatypes (e.g., Float32 and Float64), it's recommended that you use C++ and write your code using templates. Then use extern C to instantiate bindings for each datatype. For example:

template <typename T>
__device__ void kernel_function1(T *data) {
    // Code goes here
template <typename T1, typename T2>
__device__ void kernel_function2(T1 *data1, T2 *data2) {
    // Code goes here

extern "C"
    void __global__ kernel_function1_float(float *data) {kernel_function1(data);}
    void __global__ kernel_function1_double(double *data) {kernel_function1(data);}
    void __global__ kernel_function2_int_float(int *data1, float *data2) {kernel_function2(data1,data2);}

Initializing and freeing PTX modules

To easily make your kernels available, the recommended approach is to define something analogous to the following for each ptx module (this example uses the kernels described in the previous section):

module MyCudaModule

import CUDAdrv: CuModule, CuModuleFile, unload,
                CuFunction, cudacall
using CUDArt

export function1

const ptxdict = Dict()
const mdlist = Array{CuModule}(0)

function mdinit(devlist)
    global ptxdict
    global mdlist
    isempty(mdlist) || error("mdlist is not empty")
    for dev in devlist
        md = CuModuleFile("mycudamodule.ptx")
        ptxdict[(dev, "function1", Float32)] = CuFunction(md, "kernel_function1_float")
        ptxdict[(dev, "function1", Float64)] = CuFunction(md, "kernel_function1_double")
        ptxdict[(dev, "function2", Int32, Float32)] = CuFunction(md, "kernel_function2_int_float")

        push!(mdlist, md)

mdclose() = (for md in mdlist; unload(md); end; empty!(mdlist); empty!(ptxdict))

function init(f::Function, devlist)
    local ret
        ret = f(devlist)

function function1{T}(data::CudaArray{T})
    dev = device(data)
    cufunction1 = ptxdict[(dev, "function1", T)]
    # Set up grid and block, see below
    cudacall(cufunction1, grid, block, (Ptr{T},), data)


end  # MyCudaModule

Usage will look something like the following:

using CUDArt, MyCudaModule

A = rand(10,5)

result = devices(dev->capability(dev)[1]>=2) do devlist
    MyCudaModule.init(devlist) do dev

Grid and block dimensions

To be written.


One can use streams to manage or synchronize computations between the CPU & GPU, or using multiple CUDA devices. Using Julia's @sync and @async macros, here is a short demonstration that activates processing on multiple devices:

measured_sleep_time = CUDArt.devices(dev->true, nmax=2) do devlist
    sleeptime = 0.5
    results = Array{Float64}(3*length(devlist))
    streams = [(device(dev); Stream()) for dev in devlist]
    # Force one run to precompile
    cudasleep(sleeptime; dev=devlist[1], stream=streams[1])
    i = 1
    nextidx() = (idx=i; i+=1; idx)
    @sync begin
        for idev = 1:length(devlist)
            @async begin
                while true
                    idx = nextidx()
                    if idx > length(results)
                    tstart = time()
                    dev = devlist[idev]
                    stream = streams[idev]
                    cudasleep(sleeptime; dev=dev, stream=stream)
                    tstop = time()
                    results[idx] = tstop-tstart

In a more realistic version of this demonstration, you would "feed" work and collect the results from all of your CUDA devices using a single Julia process to organize the efforts.

Random notes

Notes on memory

Julia convention is that matrices are stored in column-major order, whereas C (and CUDA) use row-major. For efficiency this wrapper avoids reordering memory, so that the linear sequence of addresses is the same between main memory and the GPU. For most usages, this is probably what you want.

However, for the purposes of linear algebra, this effectively means that one is storing the transpose of matrices on the GPU. (TODO: create CudaMatrix and CudaPitchedMatrix types that will automatically take the transpose when copying between main and GPU memory. This will be useful for cuBLAS.)

Note that the size of a CudaArray/CudaPitchedArray is represented as the size of the corresponding main-memory object; thus, an array's dimensions (as reported by Julia) will not change when you copy it between main and GPU memory.


First Commit


Last Touched

about 1 month ago


196 commits