Bogdan Opanchuk <bog...@opanchuk.net> writes: > Consider the following code: > > import numpy > import pyopencl as cl > import pyopencl.array as array > > def to_device(ctx, queue, arr): > size = arr.size * arr.dtype.itemsize > buf = cl.Buffer(ctx, cl.mem_flags.READ_WRITE, size=size) > arr_dev = array.Array(queue, arr.shape, arr.dtype, data=buf) > arr_dev.set(arr, queue=queue, async_=True) > return arr_dev > > ctx = cl.create_some_context() > queue = cl.CommandQueue(ctx) > > a = numpy.empty(1024 * 4000, numpy.uint64) > ad = to_device(ctx, queue, a) > > b = numpy.empty(1024 * 4000, numpy.int32) > bd = to_device(ctx, queue, b) > > c = numpy.empty(1024 * 32, numpy.int32) > cd = to_device(ctx, queue, c) > > # queue.finish() # uncommenting this line fixes the problem > > ad.get() > > When I run it on Linux, Tesla P100 and using CUDA as the OpenCL platform, > most of the time (not always) the execution hangs on the last line. Does > anyone have any ideas about what might be happening here? > > (If I just use array.to_device(), the problem disappears. This is an > extract from a larger code, where a separate creation of the buffer is > necessary)
That code looks completely innocent to me--no idea what should be wrong with it. I just tried it on a TITAN V with the 396.54 drivers, and it ran OK without hanging for at least 40 runs. Can you get a backtrace of the hang? (Maybe attach gdb while it's hanging?) My first suspicion would be an Nv driver issue. Andreas
signature.asc
Description: PGP signature
_______________________________________________ PyOpenCL mailing list -- pyopencl@tiker.net To unsubscribe send an email to pyopencl-le...@tiker.net
