def test_callback(self):
self.callback_called = False
self.py_event = PyEvent()
def callback(event, status):
self.callback_called = True
self.py_event.set()
event = UserEvent(ctx)
queue = Queue(ctx, ctx.devices[0])
queue.enqueue_wait_for_events(event)
event2 = queue.marker()
event2.add_callback(callback)
self.assertEqual(event.status, Event.SUBMITTED)
self.assertEqual(event2.status, Event.QUEUED)
self.assertFalse(self.callback_called)
event.complete()
self.assertEqual(event.status, Event.COMPLETE)
event2.wait()
self.assertEqual(event2.status, Event.COMPLETE)
event_is_set = self.py_event.wait(2)
self.assertTrue(event_is_set, 'timed out waiting for callback')
self.assertTrue(self.callback_called)
def test_copy_contig(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
copy_of = clbuf.copy(queue)
queue.barrier()
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
self.assertTrue(np.all(a == b))
def test_copy_contig(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
copy_of = clbuf.copy(queue)
queue.barrier()
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
self.assertTrue(np.all(a == b))
def test_dim_reduce(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([[1, 2], [3, 4], [5, 6]])
view = DeviceMemoryView.from_host(ctx, a)
new_view = view[:, 0]
self.assertEqual(new_view.ndim, a[:, 0].ndim)
self.assertEqual(new_view.shape, a[:, 0].shape)
self.assertEqual(new_view.offset_, 0)
self.assertEqual(new_view.strides, a[:, 0].strides)
with new_view.map(queue) as buf:
b = np.asarray(buf)
self.assertTrue(np.all(b == a[:, 0]))
new_view = view[:, 1]
with new_view.map(queue) as buf:
b = np.asarray(buf)
self.assertTrue(np.all(b == a[:, 1]))
new_view = view[0, :]
with new_view.map(queue) as buf:
b = np.asarray(buf)
self.assertTrue(np.all(b == a[0, :]))
new_view = view[1, :]
with new_view.map(queue) as buf:
b = np.asarray(buf)
self.assertTrue(np.all(b == a[1, :]))
def test_copy_2D(self):
queue = Queue(ctx, ctx.devices[0])
a = np.arange(6 * 6).reshape([6, 6])
clbuf = DeviceMemoryView.from_host(ctx, a)
slices = [
(slice(None, None, 2), slice(None, None, 2)),
(slice(1, None, None), slice(1, None, None)),
(slice(None, None, None), slice(1, None, None)),
(slice(1, None, None), slice(None, None, None)),
(slice(1, None, None), slice(0, None, 2)),
(slice(None, None, 2), slice(1, None, 2)),
(slice(1, None, 2), slice(None, None, 2)),
]
for idx0, idx1 in slices:
expected = a[idx0, idx1]
sub_buf = clbuf[idx0, idx1]
copy_of = sub_buf.copy(queue)
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
expected = a[idx0, idx1]
self.assertTrue(np.all(expected == b), (idx0, idx1))
def test_get_slice(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([1, 2, 3, 4])
clbuf = DeviceMemoryView.from_host(ctx, a)
self.assertEqual(clbuf._refcount, 1)
new_buf = clbuf[::2]
with new_buf.map(queue) as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[::2]))
new_buf = clbuf[1::2]
with new_buf.map(queue) as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[1::2]))
new_buf = clbuf[::-1]
with new_buf.map(queue) as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[::-1]))
def test_copy_1D(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([1, 2, 3, 4])
clbuf = DeviceMemoryView.from_host(ctx, a)
copy_of = clbuf[::2].copy(queue)
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
self.assertTrue(np.all(a[::2] == b))
copy_of = clbuf[1::2].copy(queue)
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
self.assertTrue(np.all(a[1::2] == b))
copy_of = clbuf[1:-1].copy(queue)
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
self.assertTrue(np.all(a[1:-1] == b))
def test_refcount(self):
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
self.assertEqual(clbuf._refcount, 1)
new_buf = clbuf[:, :-1]
self.assertEqual(clbuf._refcount, 2)
del new_buf
gc.collect()
self.assertEqual(clbuf._refcount, 1)
self.assertEqual(clbuf.base, None)
#create sub_buffer
new_buf = clbuf[1, :]
self.assertEqual(clbuf._refcount, 2)
del new_buf
gc.collect()
self.assertEqual(clbuf._refcount, 1)
queue = Queue(ctx)
with clbuf.map(queue) as host:
self.assertEqual(clbuf._refcount, 1)
self.assertEqual(clbuf._refcount, 2, "unmap increments the refcount")
del host
gc.collect()
#GPU may not decrement the ref count
#unless finish is called
queue.finish()
self.assertEqual(clbuf._refcount, 1)
event = PyEvent()
def callback(mem):
event.set()
def test_refcount(self):
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
self.assertEqual(clbuf._refcount, 1)
new_buf = clbuf[:, :-1]
self.assertEqual(clbuf._refcount, 2)
del new_buf
gc.collect()
self.assertEqual(clbuf._refcount, 1)
self.assertEqual(clbuf.base, None)
#create sub_buffer
new_buf = clbuf[1, :]
self.assertEqual(clbuf._refcount, 2)
del new_buf
gc.collect()
self.assertEqual(clbuf._refcount, 1)
queue = Queue(ctx)
with clbuf.map(queue) as host:
self.assertEqual(clbuf._refcount, 1)
self.assertEqual(clbuf._refcount, 2, "unmap increments the refcount")
del host
gc.collect()
#GPU may not decrement the ref count
#unless finish is called
queue.finish()
self.assertEqual(clbuf._refcount, 1)
event = PyEvent()
def callback(mem):
event.set()
def test_wait(self):
event = UserEvent(ctx)
queue = Queue(ctx, ctx.devices[0])
queue.enqueue_wait_for_events(event)
event2 = queue.marker()
self.assertEqual(event.status, Event.SUBMITTED)
self.assertEqual(event2.status, Event.QUEUED)
event.complete()
self.assertEqual(event.status, Event.COMPLETE)
event2.wait()
self.assertEqual(event2.status, Event.COMPLETE)
def test_enqueue_native_kernel(self):
if not ctx.devices[0].has_native_kernel:
self.skipTest("Device does not support native kernels")
queue = Queue(ctx, ctx.devices[0])
global foo
foo = 0
def incfoo(arg, op=lambda a, b: 0):
global foo
foo = op(foo, arg)
queue.enqueue_native_kernel(incfoo, 4, op=lambda a, b: a + b)
queue.enqueue_native_kernel(incfoo, 3, op=lambda a, b: a * b)
queue.finish()
self.assertEqual(foo, 12)
def test_read_write(self):
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
queue = Queue(ctx, ctx.devices[0])
out = np.zeros_like(a)
clbuf.read(queue, out, blocking=True)
self.assertTrue(np.all(out == a))
clbuf.write(queue, a + 1, blocking=True)
clbuf.read(queue, out, blocking=True)
self.assertTrue(np.all(out == a + 1))
def test_map(self):
image_format = cl.ImageFormat('CL_RGBA', 'CL_UNSIGNED_INT8')
image = cl.empty_image(ctx, [4, 4], image_format)
queue = Queue(ctx)
with image.map(queue) as img:
self.assertEqual(img.format, 'T{B:r:B:g:B:b:B:a:}')
self.assertEqual(img.ndim, 2)
self.assertEqual(img.shape, (4, 4))
array = np.asarray(img)
array['r'] = 1
with image.map(queue) as img:
array = np.asarray(img)
self.assertTrue(np.all(array['r'] == 1))
def test_set_args(self):
program = Program(ctx, source=source)
program.build()
generate_sin = program.kernel('generate_sin')
generate_sin.argtypes = [global_memory(), ctypes.c_float]
buf = empty(ctx, [10], ctype=cl.cl_float2)
queue = Queue(ctx, ctx.devices[0])
generate_sin.set_args(buf, 1.0)
queue.enqueue_nd_range_kernel(generate_sin, 1, global_work_size=[buf.size])
expected = np.zeros([10], dtype=[('x', np.float32), ('y', np.float32)])
expected['x'] = np.arange(10)
expected['y'] = np.sin(expected['x'] / 10)
with buf.map(queue) as host:
self.assertTrue(np.all(expected['x'] == np.asarray(host)[:, 0]))
self.assertTrue(np.allclose(expected['y'], np.asarray(host)[:, 1]))
generate_sin.argnames = ['a', 'scale']
generate_sin.set_args(a=buf, scale=1.0)
queue.enqueue_nd_range_kernel(generate_sin, 1, global_work_size=[buf.size])
with buf.map(queue) as host:
self.assertTrue(np.all(expected['x'] == np.asarray(host)[:, 0]))
self.assertTrue(np.allclose(expected['y'], np.asarray(host)[:, 1]))
with self.assertRaises(TypeError):
generate_sin.set_args(a=buf)
generate_sin.__defaults__ = [1.0]
generate_sin.set_args(a=buf)
queue.enqueue_nd_range_kernel(generate_sin, 1, global_work_size=[buf.size])
with buf.map(queue) as host:
self.assertTrue(np.all(expected['x'] == np.asarray(host)[:, 0]))
self.assertTrue(np.allclose(expected['y'], np.asarray(host)[:, 1]))
def test_map(self):
a = np.array([[1, 2], [3, 4]])
view_a = memoryview(a)
clbuf = DeviceMemoryView.from_host(ctx, a)
queue = Queue(ctx, ctx.devices[0])
self.assertEqual(clbuf._mapcount, 0)
with clbuf.map(queue, writeable=False) as buf:
self.assertEqual(clbuf._mapcount, 1)
self.assertEqual(buf.format, view_a.format)
self.assertEqual(buf.shape, view_a.shape)
self.assertEqual(buf.strides, view_a.strides)
b = np.asarray(buf)
self.assertTrue(np.all(b == a))
self.assertTrue(buf.readonly)
# self.assertEqual(clbuf._mapcount, 0)
with clbuf.map(queue, readable=False) as buf:
self.assertEqual(clbuf._mapcount, 1)
b = np.asarray(buf)
b[::] = a[::-1]
self.assertFalse(buf.readonly)
# self.assertEqual(clbuf._mapcount, 0)
with clbuf.map(queue, writeable=False) as buf:
self.assertEqual(clbuf._mapcount, 1)
b = np.asarray(buf)
self.assertTrue(np.all(b == a[::-1]))
def test_getitem(self):
queue = Queue(ctx, ctx.devices[0])
a = np.array([[1, 2], [3, 4]])
clbuf = DeviceMemoryView.from_host(ctx, a)
with self.assertRaises(IndexError):
clbuf[1, 1, 1]
self.assertEqual(clbuf._refcount, 1)
new_buf = clbuf[:, :-1]
self.assertEqual(clbuf._refcount, 2)
mapp = new_buf.map(queue)
with mapp as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[:, :-1]))
del buf
del new_buf
gc.collect()
new_buf = clbuf[:, 1:]
with new_buf.map(queue) as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[:, 1:]))
new_buf = clbuf[1:, :]
with new_buf.map(queue) as buf:
b = np.asanyarray(buf)
self.assertTrue(np.all(b == a[1:, :]))
def test_enqueue_native_kernel_refcount(self):
if not ctx.devices[0].has_native_kernel:
self.skipTest("Device does not support native kernels")
queue = Queue(ctx, ctx.devices[0])
def incfoo():
pass
self.assertEqual(sys.getrefcount(incfoo), 2)
e = cl.UserEvent(ctx)
queue.enqueue_wait_for_events(e)
queue.enqueue_native_kernel(incfoo)
self.assertEqual(sys.getrefcount(incfoo), 3)
e.complete()
queue.finish()
self.assertEqual(sys.getrefcount(incfoo), 2)
def test_copy_2D_negative_stride(self):
queue = Queue(ctx, ctx.devices[0])
a = np.arange(4 * 4).reshape([4, 4])
clbuf = DeviceMemoryView.from_host(ctx, a)
slices = [
(slice(None, None, -2), slice(None, None, -2)),
(slice(1, None, -1), slice(1, None, -1)),
(slice(None, None, None), slice(1, None, -1)),
(slice(1, None, -1), slice(None, None, -1)),
(slice(1, None, -2), slice(1, None, -2)),
(slice(None, None, -2), slice(1, None, -2)),
(slice(1, None, -2), slice(None, None, -2)),
]
for idx0, idx1 in slices:
copy_of = clbuf[idx0, idx1].copy(queue)
with copy_of.map(queue) as cpy:
b = np.asarray(cpy)
expected = a[idx0, idx1]
self.assertTrue(np.all(expected == b))