Interprocess communication with pyzmq and multiprocessing

pycuda_req_rep_demo.py

#!/usr/bin/env python

"""
Pass data between processes started through the multiprocessing module
using pyzmq and process them with PyCUDA
"""

import numpy as np
import zmq
import multiprocessing as mp

gpu = 0

def worker():
    import pycuda.driver as drv
    import pycuda.gpuarray as gpuarray
    import atexit

    # Initialize device:
    drv.init()
    dev = drv.Device(gpu)
    ctx = dev.make_context()
    atexit.register(ctx.pop)
    
    print dev.name()
       
    context = zmq.Context()
    socket = context.socket(zmq.REP)
    socket.connect("tcp://localhost:5555")

    # Process data sent to worker until a quit signal is transmitted:
    while True:
        data = socket.recv_pyobj()
        print "Worker %i: %s" % (gpu, data)
        if data == 'quit':
            break

        # Do something with the data on the GPU:
        data_gpu = gpuarray.to_gpu(data)
        result_gpu = -data_gpu

        socket.send_pyobj(result_gpu.get())

def master():

    # Data to send to worker:
    data_list = map(lambda x: np.random.rand(4, 4), xrange(4))

    context = zmq.Context()
    socket = context.socket(zmq.REQ)
    socket.bind("tcp://*:5555")

    # Send data out for processing and get back the results:
    for i in xrange(len(data_list)):
        socket.send_pyobj(data_list[i])
        result = socket.recv_pyobj()
        print "Master: ", result
    socket.send_pyobj('quit') 

if __name__ == '__main__':    
    worker = mp.Process(target=worker)
    worker.start()

    master()