def test_empty_tensor_to_host(self):
buf = DeviceBuffer(shape=(5, 2, 0, 3, 0), dtype=np.float32)
assert misc.volume(buf.shape) == 0
host_buf = np.empty(tuple(), dtype=np.float32)
assert misc.volume(host_buf.shape) == 1
host_buf = buf.copy_to(host_buf)
assert host_buf.shape == buf.shape
assert host_buf.nbytes == 0
assert misc.volume(host_buf.shape) == 0
def copy_to(self, host_buffer, stream=None):
"""
Copies from this device buffer to the provided host buffer.
Host buffer must be contiguous in memory (see np.ascontiguousarray).
Args:
host_buffer (numpy.ndarray): The host buffer to copy into.
stream (Stream):
A Stream instance (see util/cuda.py). Performs a synchronous copy if no stream is provided.
Returns:
numpy.ndarray: The host buffer, possibly reallocated if the provided buffer was too small.
"""
nbytes = misc.volume(self.shape) * np.dtype(self.dtype).itemsize
self._check_dtype_matches(host_buffer)
try:
host_buffer.resize(self.shape, refcheck=False)
except ValueError:
host_buffer = np.empty(self.shape, dtype=np.dtype(self.dtype))
if nbytes:
host_ptr = host_buffer.ctypes.data_as(ctypes.c_void_p)
wrapper().dtoh(dst=host_ptr,
src=self._ptr,
nbytes=nbytes,
stream=try_get_stream_handle(stream))
host_buffer = host_buffer.reshape(self.shape)
return host_buffer
def check_empty_tensor_expand(runner, shapes):
shape = shapes["new_shape"]
feed_dict = {
"data": np.array((0, )),
"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 misc.volume(outputs["expanded"].shape) == 0
def resize(self, shape):
nbytes = misc.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 misc.volume(it) == vol
