Study impact of syncrhonizing before host-device and device-host memory copies
With MR !458 (merged) host memory is no longer freed within a sequence of algorithms to avoid problems that can arise from host memory having changed before copies to / from it have happened due to the asynchronous execution on the device.
An alternative way to avoid this is to synchronize the stream before host-device and device-host copies. This can however lead to significant performance loss and its impact should be studied.