Remove C++ Wrappers in memory_resource_adaptors.hpp Needed by Cython (

#662)

Depends on #661 

While working on PR #661, it looked like it was possible to remove the "owning" C++ memory resource adaptors in `memory_resource_adaptors.hpp`. This PR is a quick implementation of what that would look like to run through CI and get feedback.

The main driving factor of this PR is to eliminate the need for 2 layers of wrappers around every memory resource in the library. When adding new memory resources, C++ wrappers must be created in `memory_resource_adaptors.hpp` and Cython wrappers must be created in `memory_resource.pyx`, for any property/function that needs to be exposed at the python level. This removes the C++ wrappers in favor of using pythons reference counting for lifetime management.

A few notes:

1. `MemoryResource` was renamed `DeviceMemoryResource` to more closely match the C++ class names. Easily can be changed back
1. Upstream MR are kept alive by a base class `UpstreamResourceAdaptor` that stores a single property `upstream_mr`. Any MR that has an upstream, needs to inherit from this class.
1. Once the `UpstreamResourceAdaptor` was created, most of the work/changes were updating the Cython imports to use the C++ classes instead of the C++ wrappers.
1. This should make it easier to expose more methods/properties at the python layer in the future.

Would appreciate any feedback.

Authors:
  - Michael Demoret (@mdemoret-nv)

Approvers:
  - Christopher Harris (@cwharris)
  - Keith Kraus (@kkraus14)
  - Ashwin Srinath (@shwina)

URL: #662

python/rmm/_lib/device_buffer.pxd

@@ -12,12 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from libcpp.memory cimport unique_ptr
 from libc.stdint cimport uintptr_t
+from libcpp.memory cimport unique_ptr
 
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
 from rmm._cuda.stream cimport Stream
-from rmm._lib.memory_resource cimport MemoryResource
+from rmm._lib.cuda_stream_view cimport cuda_stream_view
+from rmm._lib.memory_resource cimport DeviceMemoryResource
 
 
 cdef extern from "rmm/device_buffer.hpp" namespace "rmm" nogil:
@@ -39,10 +39,10 @@ cdef extern from "rmm/device_buffer.hpp" namespace "rmm" nogil:
 cdef class DeviceBuffer:
     cdef unique_ptr[device_buffer] c_obj
 
-    # Holds a reference to the MemoryResource used for allocation. Ensures the
-    # MR does not get destroyed before this DeviceBuffer. `mr` is needed for
-    # deallocation
-    cdef MemoryResource mr
+    # Holds a reference to the DeviceMemoryResource used for allocation.
+    # Ensures the MR does not get destroyed before this DeviceBuffer. `mr` is
+    # needed for deallocation
+    cdef DeviceMemoryResource mr
 
     # Holds a reference to the stream used by the underlying `device_buffer`.
     # Ensures the stream does not get destroyed before this DeviceBuffer

python/rmm/_lib/memory_resource.pxd

@@ -1,108 +1,50 @@
 # Copyright (c) 2020, NVIDIA CORPORATION.
 
 from libc.stdint cimport int8_t
-from libcpp.vector cimport vector
-from libcpp.string cimport string
 from libcpp.memory cimport shared_ptr
+from libcpp.string cimport string
+from libcpp.vector cimport vector
+
+
+cdef extern from "rmm/mr/device/device_memory_resource.hpp" \
+        namespace "rmm::mr" nogil:
+    cdef cppclass device_memory_resource:
+        pass
 
+cdef class DeviceMemoryResource:
+    cdef shared_ptr[device_memory_resource] c_obj
 
-cdef extern from "memory_resource_wrappers.hpp" nogil:
-    cdef cppclass device_memory_resource_wrapper:
-        shared_ptr[device_memory_resource_wrapper] get_mr() except +
-
-    cdef cppclass default_memory_resource_wrapper(
-        device_memory_resource_wrapper
-    ):
-        default_memory_resource_wrapper(int device) except +
-
-    cdef cppclass cuda_memory_resource_wrapper(device_memory_resource_wrapper):
-        cuda_memory_resource_wrapper() except +
-
-    cdef cppclass managed_memory_resource_wrapper(
-        device_memory_resource_wrapper
-    ):
-        managed_memory_resource_wrapper() except +
-
-    cdef cppclass pool_memory_resource_wrapper(device_memory_resource_wrapper):
-        pool_memory_resource_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr,
-            size_t initial_pool_size,
-            size_t maximum_pool_size
-        ) except +
-
-    cdef cppclass fixed_size_memory_resource_wrapper(
-        device_memory_resource_wrapper
-    ):
-        fixed_size_memory_resource_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr,
-            size_t block_size,
-            size_t blocks_to_preallocate
-        ) except +
-
-    cdef cppclass binning_memory_resource_wrapper(
-        device_memory_resource_wrapper
-    ):
-        binning_memory_resource_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr
-        ) except +
-        binning_memory_resource_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr,
-            int8_t min_size_exponent,
-            int8_t max_size_exponent
-        ) except +
-        void add_bin(
-            size_t allocation_size,
-            shared_ptr[device_memory_resource_wrapper] bin_mr
-        ) except +
-        void add_bin(
-            size_t allocation_size
-        ) except +
-
-    cdef cppclass logging_resource_adaptor_wrapper(
-        device_memory_resource_wrapper
-    ):
-        logging_resource_adaptor_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr,
-            string filename
-        ) except +
-        void flush() except +
-
-    cdef cppclass thread_safe_resource_adaptor_wrapper(
-        device_memory_resource_wrapper
-    ):
-        thread_safe_resource_adaptor_wrapper(
-            shared_ptr[device_memory_resource_wrapper] upstream_mr,
-        ) except +
-
-    void set_per_device_resource(
-        int device,
-        shared_ptr[device_memory_resource_wrapper] new_resource
-    ) except +
-
-
-cdef class MemoryResource:
-    cdef shared_ptr[device_memory_resource_wrapper] c_obj
-
-cdef class CudaMemoryResource(MemoryResource):
+    cdef device_memory_resource* get_mr(self)
+
+cdef class UpstreamResourceAdaptor(DeviceMemoryResource):
+    cdef readonly DeviceMemoryResource upstream_mr
+
+    cpdef DeviceMemoryResource get_upstream(self)
+
+cdef class CudaMemoryResource(DeviceMemoryResource):
     pass
 
-cdef class ManagedMemoryResource(MemoryResource):
+cdef class ManagedMemoryResource(DeviceMemoryResource):
     pass
 
-cdef class PoolMemoryResource(MemoryResource):
+cdef class PoolMemoryResource(UpstreamResourceAdaptor):
     pass
 
-cdef class FixedSizeMemoryResource(MemoryResource):
+cdef class FixedSizeMemoryResource(UpstreamResourceAdaptor):
     pass
 
-cdef class BinningMemoryResource(MemoryResource):
-    cpdef add_bin(self, size_t allocation_size, object bin_resource=*)
+cdef class BinningMemoryResource(UpstreamResourceAdaptor):
+
+    cdef readonly list bin_mrs
+
+    cpdef add_bin(
+        self,
+        size_t allocation_size,
+        DeviceMemoryResource bin_resource=*)
 
-cdef class LoggingResourceAdaptor(MemoryResource):
-    cdef MemoryResource upstream
+cdef class LoggingResourceAdaptor(UpstreamResourceAdaptor):
     cdef object _log_file_name
-    cpdef MemoryResource get_upstream(self)
     cpdef get_file_name(self)
     cpdef flush(self)
 
-cpdef MemoryResource get_current_device_resource()
+cpdef DeviceMemoryResource get_current_device_resource()