Integrate OpenCL <-> Julia Event systems

[jakebolewski]

We should have better integration with Julia's underlying event / task system.

[vchuravy]

My intuition would be that with http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clSetEventCallback.html we should be able to build either RemoteReferences or Tasks that completes when the underlying events complete.

Also of interest might be something like https://github.com/shashi/React.jl

[vchuravy]

Something along those lines might be helpful. More information here: http://julialang.org/blog/2013/05/callback/

using OpenCL; const cl = OpenCL

type Event
    status :: Symbol
    error_code :: Int
    Event() = Event(:undefined)
    Event(status :: Symbol) = new(status, 0)
end 

device = first(cl.devices())
ctx = cl.Context(device)
q = cl.CmdQueue(ctx)

test = Event()

function callback(event, event_status, data :: Ptr{Void})
    event_data = unsafe_pointer_to_objref(data) :: Event

    error_code = 0
    if event_status == cl.CL_COMPLETE
        status = :complete
    elseif event_status == cl.CL_SUBMITTED
        status = :submitted
    elseif event_status == cl.CL_RUNNING
        status = :submitted
    elseif event_status == cl.CL_QUEUED
        status = :queued
    elseif event_status < 0
        status = :error
        error_code = event_status
    end

    event_data.status = status
    event_data.error_code = error_code

    return nothing
end


const callback_c = cfunction(callback, Void, (cl.CL_event, cl.CL_int, Ptr{Void}))

usr_evt = cl.UserEvent(ctx)
cl.enqueue_wait_for_events(q, usr_evt)
mkr_evt = cl.enqueue_marker(q)

cl.api.clSetEventCallback(mkr_evt.id, cl.CL_COMPLETE, callback_c, test)
println(test)
cl.complete(usr_evt)
sleep(1.0)
println(test)

[jakebolewski]

@vchuravy that is a good sketch of what we will have to do (what you are missing is a global ObjectID Dict to hold the references to the underlying Julia objects you are hiding from the GC). We also need a Condition object as a field of the Event type. This is how you signal to Julia's event loop that something has changed and that dependent code needs to be scheduled to be executed (grep Condition to see how it is used in base). Making this more difficult is the fact that callbacks can be executed asynchronously by the OpenCL runtime so we have to factor in thread safety ( http://docs.julialang.org/en/latest/manual/calling-c-and-fortran-code/#thread-safety). I don't know what the performance hit of this is so this is an abstraction that probably needs to be opted into instead of being the default.

[jakebolewski]

With CPU OpenCL implementations (and Intel's PHI) the Task parallel model that OpenCL provides is a great way to execute multiple different kernels concurrently. It is an underutilized part of the spec because it makes no sense for GPU's (currently).