def test_empty_tensor_to_host(self):
with DeviceArray(shape=(5, 2, 0, 3, 0), dtype=np.float32) as buf:
assert util.volume(buf.shape) == 0
host_buf = np.empty(tuple(), dtype=np.float32)
assert util.volume(host_buf.shape) == 1
host_buf = buf.copy_to(host_buf)
assert host_buf.shape == buf.shape
assert host_buf.nbytes == 0
assert util.volume(host_buf.shape) == 0
def test_large_allocation(self):
dtype = np.byte
# See if we can alloc 3GB (bigger than value of signed int)
shape = (3 * 1024 * 1024 * 1024, )
with DeviceArray(shape=shape, dtype=dtype) as buf:
assert buf.allocated_nbytes == util.volume(shape) * np.dtype(
dtype).itemsize
def check_empty_tensor_expand(runner, shapes):
shape = shapes["new_shape"]
feed_dict = {
"data": np.zeros(shape=(2, 0, 3, 0), dtype=np.float32),
"new_shape": np.array(shape, dtype=np.int32)
}
outputs = runner.infer(feed_dict)
# Empty tensor will still be empty after broadcast
assert outputs["expanded"].shape == shape
assert util.volume(outputs["expanded"].shape) == 0
def resize(self, shape):
"""
Resizes or reshapes the array to the specified shape.
If the allocated memory region is already large enough,
no reallocation is performed.
Args:
shape (Tuple[int]): The new shape.
"""
nbytes = util.volume(shape) * np.dtype(self.dtype).itemsize
if nbytes > self.allocated_nbytes:
self.free()
self.allocate(nbytes)
self.shape = shape
def test_volume(case):
it, vol = case
assert util.volume(it) == vol
def nbytes(self):
"""
The number of bytes in the memory region.
"""
return util.volume(self.shape) * np.dtype(self.dtype).itemsize
def arange(shape):
return np.arange(util.volume(shape)).reshape(shape)