Compare USM with RMM

[PokhodenkoSA]

Rapids Memory Manager (RMM):
https://github.com/rapidsai/rmm#using-rmm-in-python-code
http://numba.pydata.org/numba-doc/latest/cuda/external-memory.html#setting-the-emm-plugin

Related to #34

[oleksandr-pavlyk]

RMM implements DeviceBuffer which is Cython native class wrapping around something similar to std::vector<unsigned char, custom_cuda_allocator (calls resource manager)> which is called device_buffer.

DeviceBuffer stores a unique pointer to an instance of this class. DeviceBuffer implements __cuda_array_interface__. Direct constructors always allocate new memory and copy provided inputs into the newly allocated array.

Zero-copy construction is possible from a unique_ptr<device_ buffer>, with the ownership being moved to the Cython extension instance.

DeviceBuffer provides __reduce__ method to support pickling (which works by copying content of the device buffer to host) and provides the following set of routines, among others:

• copy_to_host(host_buf_obj) to copy content of the underlying device_buffer to a host buffer
• copy_from_host(host_buf_obf) to copy content of the host buffer into memory of underlying device_buffer
• copy_from_device(cuda_ary_obj) to copy device memory underyling cuda_ary_obj Python object implementing __cuda_array_interface__ to the memory underlying DeviceBuffer instance.

RMM's methods are declared nogil.

[diptorupd]

I have documented the current discussion in our docs under dpctl.memory. Closing this issue.

[PokhodenkoSA]

Related: #140