[Improvement] Reorganize Cython to separate C++ bindings and make Cyt…

…hon classes public (#1676)

Closes #1280

Authors:
  - Matthew Murray (https://github.com/Matt711)
  - Lawrence Mitchell (https://github.com/wence-)

Approvers:
  - Lawrence Mitchell (https://github.com/wence-)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #1676

.gitignore

@@ -22,10 +22,13 @@ rmm.egg-info/
 python/build
 python/*/build
 python/rmm/docs/_build
-python/rmm/**/_lib/**/*.cpp
-!python/rmm/_lib/_torch_allocator.cpp
-python/rmm/**/_lib/**/*.h
-python/rmm/**/_lib/.nfs*
+python/rmm/**/librmmm/**/*.cpp
+!python/rmm/librmmm/_torch_allocator.cpp
+python/rmm/**/librmm/**/*.h
+python/rmm/**/librmm/.nfs*
+python/rmm/**/pylibrmmm/**/*.cpp
+python/rmm/**/pylibrmmm/**/*.h
+python/rmm/**/pylibrmmm/.nfs*
 python/rmm/_cuda/*.cpp
 python/rmm/tests/*.cpp
 python/rmm/*.ipynb

python/rmm/CMakeLists.txt

@@ -30,4 +30,5 @@ rapids_cython_init()
 add_compile_definitions("SPDLOG_ACTIVE_LEVEL=SPDLOG_LEVEL_${RMM_LOGGING_LEVEL}")
 
 add_subdirectory(rmm/_cuda)
-add_subdirectory(rmm/_lib)
+add_subdirectory(rmm/librmm)
+add_subdirectory(rmm/pylibrmm)

python/rmm/docs/guide.md

@@ -236,17 +236,17 @@ Common to both usages is that they modify the currently active RMM memory resour
 
 >>> # We start with the default cuda memory resource
 >>> rmm.mr.get_current_device_resource()
-<rmm._lib.memory_resource.CudaMemoryResource at 0x7f7e6c0a1ce0>
+<rmm.pylibrmm.memory_resource.CudaMemoryResource object at 0x7fa0da48a8e0>
 
 >>> # When using statistics, we get a StatisticsResourceAdaptor with the context
 >>> with rmm.statistics.statistics():
 ...     rmm.mr.get_current_device_resource()
-<rmm._lib.memory_resource.StatisticsResourceAdaptor at 0x7f7e6c524900>
+<rmm.pylibrmm.memory_resource.StatisticsResourceAdaptor object at 0x7fa0dd6e4a40>
 
 >>> # We can also enable statistics globally
 >>> rmm.statistics.enable_statistics()
 >>> print(rmm.mr.get_current_device_resource())
-<rmm._lib.memory_resource.StatisticsResourceAdaptor at 0x7f662c2bb3c0>
+<rmm.pylibrmm.memory_resource.StatisticsResourceAdaptor object at 0x7f9a11340a40>
 ```
 
 With statistics enabled, you can query statistics of the current and peak bytes and number of allocations performed by the current RMM memory resource:

python/rmm/rmm/__init__.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2018-2021, NVIDIA CORPORATION.
+# Copyright (c) 2018-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -13,8 +13,10 @@
 # limitations under the License.
 
 from rmm import mr
-from rmm._lib.device_buffer import DeviceBuffer
-from rmm._lib.logger import (
+from rmm._version import __git_commit__, __version__
+from rmm.mr import disable_logging, enable_logging, get_log_filenames
+from rmm.pylibrmm.device_buffer import DeviceBuffer
+from rmm.pylibrmm.logger import (
     flush_logger,
     get_flush_level,
     get_logging_level,
@@ -23,8 +25,6 @@
     set_logging_level,
     should_log,
 )
-from rmm._version import __git_commit__, __version__
-from rmm.mr import disable_logging, enable_logging, get_log_filenames
 from rmm.rmm import (
     RMMError,
     is_initialized,
@@ -52,3 +52,13 @@
     "should_log",
     "unregister_reinitialize_hook",
 ]
+
+
+def __getattr__(name):
+    if name == "_lib":
+        import importlib
+
+        module = importlib.import_module("rmm.pylibrmm")
+        return module
+    else:
+        raise AttributeError(f"Module '{__name__}' has no attribute '{name}'")

python/rmm/rmm/_cuda/stream.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2020, NVIDIA CORPORATION.
+# Copyright (c) 2020-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@ from cuda.ccudart cimport cudaStream_t
 from libc.stdint cimport uintptr_t
 from libcpp cimport bool
 
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
+from rmm.librmm.cuda_stream_view cimport cuda_stream_view
 
 
 cdef class Stream:

python/rmm/rmm/_cuda/stream.pyx

@@ -16,13 +16,13 @@ from cuda.ccudart cimport cudaStream_t
 from libc.stdint cimport uintptr_t
 from libcpp cimport bool
 
-from rmm._lib.cuda_stream cimport CudaStream
-from rmm._lib.cuda_stream_view cimport (
+from rmm.librmm.cuda_stream_view cimport (
     cuda_stream_default,
     cuda_stream_legacy,
     cuda_stream_per_thread,
     cuda_stream_view,
 )
+from rmm.pylibrmm.cuda_stream cimport CudaStream
 
 
 cdef class Stream:

python/rmm/rmm/_lib/__init__.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2021, NVIDIA CORPORATION.
+# Copyright (c) 2018-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .device_buffer import DeviceBuffer
+from rmm.pylibrmm import *

python/rmm/rmm/_lib/cuda_stream.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2020, NVIDIA CORPORATION.
+# Copyright (c) 2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,26 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-cimport cython
-from cuda.ccudart cimport cudaStream_t
-from libcpp cimport bool
-from libcpp.memory cimport unique_ptr
-
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
-
-
-cdef extern from "rmm/cuda_stream.hpp" namespace "rmm" nogil:
-    cdef cppclass cuda_stream:
-        cuda_stream() except +
-        bool is_valid() except +
-        cudaStream_t value() except +
-        cuda_stream_view view() except +
-        void synchronize() except +
-        void synchronize_no_throw()
-
-
-@cython.final
-cdef class CudaStream:
-    cdef unique_ptr[cuda_stream] c_obj
-    cdef cudaStream_t value(self) except * nogil
-    cdef bool is_valid(self) except * nogil
+from rmm.librmm.cuda_stream cimport cuda_stream
+from rmm.pylibrmm.cuda_stream cimport CudaStream

python/rmm/rmm/_lib/cuda_stream_pool.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2021, NVIDIA CORPORATION.
+# Copyright (c) 2021-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,14 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-cimport cython
-
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
-
-
-cdef extern from "rmm/cuda_stream_pool.hpp" namespace "rmm" nogil:
-    cdef cppclass cuda_stream_pool:
-        cuda_stream_pool(size_t pool_size)
-        cuda_stream_view get_stream()
-        cuda_stream_view get_stream(size_t stream_id) except +
-        size_t get_pool_size()
+from rmm.librmm.cuda_stream_pool cimport cuda_stream_pool

python/rmm/rmm/_lib/cuda_stream_view.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2020, NVIDIA CORPORATION.
+# Copyright (c) 2020-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,21 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from cuda.ccudart cimport cudaStream_t
-from libcpp cimport bool
-
-
-cdef extern from "rmm/cuda_stream_view.hpp" namespace "rmm" nogil:
-    cdef cppclass cuda_stream_view:
-        cuda_stream_view()
-        cuda_stream_view(cudaStream_t)
-        cudaStream_t value()
-        bool is_default()
-        bool is_per_thread_default()
-        void synchronize() except +
-
-    cdef bool operator==(cuda_stream_view const, cuda_stream_view const)
-
-    const cuda_stream_view cuda_stream_default
-    const cuda_stream_view cuda_stream_legacy
-    const cuda_stream_view cuda_stream_per_thread
+from rmm.librmm.cuda_stream_view cimport (
+    cuda_stream_default,
+    cuda_stream_legacy,
+    cuda_stream_per_thread,
+    cuda_stream_view,
+)

python/rmm/rmm/_lib/device_buffer.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2019-2024, NVIDIA CORPORATION.
+# Copyright (c) 2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,105 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from libc.stdint cimport uintptr_t
-from libcpp.memory cimport unique_ptr
-
-from rmm._cuda.stream cimport Stream
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
-from rmm._lib.memory_resource cimport (
-    DeviceMemoryResource,
-    device_memory_resource,
+from rmm.librmm.device_buffer cimport (
+    cuda_device_id,
+    device_buffer,
+    get_current_cuda_device,
+    prefetch,
+)
+from rmm.pylibrmm.device_buffer cimport (
+    DeviceBuffer,
+    copy_device_to_ptr,
+    copy_host_to_ptr,
+    copy_ptr_to_host,
+    to_device,
 )
-
-
-cdef extern from "rmm/mr/device/per_device_resource.hpp" namespace "rmm" nogil:
-    cdef cppclass cuda_device_id:
-        ctypedef int value_type
-        cuda_device_id()
-        cuda_device_id(value_type id)
-        value_type value()
-
-    cdef cuda_device_id get_current_cuda_device()
-
-cdef extern from "rmm/prefetch.hpp" namespace "rmm" nogil:
-    cdef void prefetch(const void* ptr,
-                       size_t bytes,
-                       cuda_device_id device,
-                       cuda_stream_view stream) except +
-
-cdef extern from "rmm/device_buffer.hpp" namespace "rmm" nogil:
-    cdef cppclass device_buffer:
-        device_buffer()
-        device_buffer(
-            size_t size,
-            cuda_stream_view stream,
-            device_memory_resource *
-        ) except +
-        device_buffer(
-            const void* source_data,
-            size_t size,
-            cuda_stream_view stream,
-            device_memory_resource *
-        ) except +
-        device_buffer(
-            const device_buffer buf,
-            cuda_stream_view stream,
-            device_memory_resource *
-        ) except +
-        void reserve(size_t new_capacity, cuda_stream_view stream) except +
-        void resize(size_t new_size, cuda_stream_view stream) except +
-        void shrink_to_fit(cuda_stream_view stream) except +
-        void* data()
-        size_t size()
-        size_t capacity()
-
-
-cdef class DeviceBuffer:
-    cdef unique_ptr[device_buffer] c_obj
-
-    # 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
-    cdef Stream stream
-
-    @staticmethod
-    cdef DeviceBuffer c_from_unique_ptr(
-        unique_ptr[device_buffer] ptr,
-        Stream stream=*,
-        DeviceMemoryResource mr=*,
-    )
-
-    @staticmethod
-    cdef DeviceBuffer c_to_device(const unsigned char[::1] b,
-                                  Stream stream=*) except *
-    cpdef copy_to_host(self, ary=*, Stream stream=*)
-    cpdef copy_from_host(self, ary, Stream stream=*)
-    cpdef copy_from_device(self, cuda_ary, Stream stream=*)
-    cpdef bytes tobytes(self, Stream stream=*)
-
-    cdef size_t c_size(self) except *
-    cpdef void reserve(self, size_t new_capacity, Stream stream=*) except *
-    cpdef void resize(self, size_t new_size, Stream stream=*) except *
-    cpdef size_t capacity(self) except *
-    cdef void* c_data(self) except *
-
-    cdef device_buffer c_release(self) except *
-
-cpdef DeviceBuffer to_device(const unsigned char[::1] b,
-                             Stream stream=*)
-cpdef void copy_ptr_to_host(uintptr_t db,
-                            unsigned char[::1] hb,
-                            Stream stream=*) except *
-
-cpdef void copy_host_to_ptr(const unsigned char[::1] hb,
-                            uintptr_t db,
-                            Stream stream=*) except *
-
-cpdef void copy_device_to_ptr(uintptr_t d_src,
-                              uintptr_t d_dst,
-                              size_t count,
-                              Stream stream=*) except *

python/rmm/rmm/_lib/device_uvector.pxd

@@ -1,4 +1,4 @@
-# Copyright (c) 2021, NVIDIA CORPORATION.
+# Copyright (c) 2021-2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,28 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from rmm._lib.cuda_stream_view cimport cuda_stream_view
-from rmm._lib.device_buffer cimport device_buffer
-from rmm._lib.memory_resource cimport device_memory_resource
-
-
-cdef extern from "rmm/device_buffer.hpp" namespace "rmm" nogil:
-    cdef cppclass device_uvector[T]:
-        device_uvector(size_t size, cuda_stream_view  stream) except +
-        T* element_ptr(size_t index)
-        void set_element(size_t element_index, const T& v, cuda_stream_view s)
-        void set_element_async(
-            size_t element_index,
-            const T& v,
-            cuda_stream_view s
-        ) except +
-        T front_element(cuda_stream_view s) except +
-        T back_element(cuda_stream_view s) except +
-        void reserve(size_t new_capacity, cuda_stream_view stream) except +
-        void resize(size_t new_size, cuda_stream_view stream) except +
-        void shrink_to_fit(cuda_stream_view stream) except +
-        device_buffer release()
-        size_t capacity()
-        T* data()
-        size_t size()
-        device_memory_resource* memory_resource()
+from rmm.librmm.device_uvector cimport device_uvector

python/rmm/rmm/_lib/helper.pxd

@@ -12,5 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
-cdef object parse_bytes(object s) except *
+from rmm.pylibrmm.helper cimport parse_bytes