mem_bug.py

import numpy
import pyopencl as cl
import pyopencl.array as array

src = """
__kernel void test(__global int *a)
{
    int gid = get_global_id(0);
    a[gid] = gid;
}
"""

platform = cl.get_platforms()[0]
devices = platform.get_devices()
devices = [devices[0], devices[2]]
context = cl.Context(devices)
print("Using devices", devices)

queue1 = cl.CommandQueue(context, devices[0])
queue2 = cl.CommandQueue(context, devices[1])

program = cl.Program(context, src).build()

l = 16
dtype = numpy.dtype('int32')

a_dev = cl.Buffer(context, cl.mem_flags.READ_WRITE | cl.mem_flags.ALLOC_HOST_PTR, l * dtype.itemsize)

a_dev_1 = a_dev.get_sub_region(0, l // 2 * dtype.itemsize)
cl.enqueue_migrate_mem_objects(queue1, [a_dev_1])
a_dev_2 = a_dev.get_sub_region(l // 2 * dtype.itemsize, l // 2 * dtype.itemsize)
cl.enqueue_migrate_mem_objects(queue2, [a_dev_2])

queue1.finish()
queue2.finish()

program.test(queue1, (l // 2,), None, a_dev_1)
program.test(queue2, (l // 2,), None, a_dev_2)

queue1.finish()
queue2.finish()

a_res = numpy.empty(l, numpy.int32)
#cl.enqueue_migrate_mem_objects(queue1, [a_dev_2])
cl.enqueue_copy(queue1, a_res, a_dev, is_blocking=False)

queue1.finish()
queue2.finish()

print(a_res)