[WIP] Reorganize Cython to separate C++ bindings and make Cython clas…

…ses public

python/rmm/CMakeLists.txt

@@ -30,4 +30,6 @@ 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/_lib)
+add_subdirectory(rmm/cpp)
+add_subdirectory(rmm/python)

python/rmm/rmm/cpp/CMakeLists.txt

@@ -0,0 +1,35 @@
+# =============================================================================
+# Copyright (c) 2022-2024, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except
+# in compliance with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software distributed under the License
+# is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+# or implied. See the License for the specific language governing permissions and limitations under
+# the License.
+# =============================================================================
+
+set(cython_sources cpp_logger.pyx cpp_memory_resource.pyx)
+set(linked_libraries rmm::rmm)
+
+# Build all of the Cython targets
+rapids_cython_create_modules(SOURCE_FILES "${cython_sources}" LINKED_LIBRARIES "${linked_libraries}"
+                                                                               CXX)
+
+# mark all symbols in these Cython targets "hidden" by default, so they won't collide with symbols
+# loaded from other DSOs
+foreach(_cython_target IN LISTS RAPIDS_CYTHON_CREATED_TARGETS)
+  set_target_properties(${_cython_target} PROPERTIES C_VISIBILITY_PRESET hidden
+                                                     CXX_VISIBILITY_PRESET hidden)
+endforeach()
+
+add_library(_torch_allocator SHARED _torch_allocator.cpp)
+# Want the output to be called _torch_allocator.so
+set_target_properties(_torch_allocator PROPERTIES PREFIX "" SUFFIX ".so")
+target_link_libraries(_torch_allocator PRIVATE rmm::rmm)
+cmake_path(RELATIVE_PATH CMAKE_CURRENT_SOURCE_DIR BASE_DIRECTORY "${PROJECT_SOURCE_DIR}"
+           OUTPUT_VARIABLE _torch_allocator_location)
+install(TARGETS _torch_allocator DESTINATION "${_torch_allocator_location}")

python/rmm/rmm/cpp/__init__.pxd

@@ -0,0 +1,13 @@
+# Copyright (c) 2019-2020, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

python/rmm/rmm/cpp/_torch_allocator.cpp

@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2023-2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <rmm/cuda_device.hpp>
+#include <rmm/cuda_stream_view.hpp>
+#include <rmm/mr/device/per_device_resource.hpp>
+
+#include <cuda_runtime_api.h>
+
+// These signatures must match those required by CUDAPluggableAllocator in
+// github.com/pytorch/pytorch/blob/main/torch/csrc/cuda/CUDAPluggableAllocator.h
+// Since the loading is done at runtime via dlopen, no error checking
+// can be performed for mismatching signatures.
+
+/**
+ * @brief Allocate memory of at least \p size bytes.
+ *
+ * @throws rmm::bad_alloc When the requested allocation cannot be satisfied.
+ *
+ * @param size The number of bytes to allocate
+ * @param device The device whose memory resource one should use
+ * @param stream CUDA stream to perform allocation on
+ * @return Pointer to the newly allocated memory
+ */
+extern "C" void* allocate(std::size_t size, int device, void* stream)
+{
+  rmm::cuda_device_id const device_id{device};
+  rmm::cuda_set_device_raii with_device{device_id};
+  auto mr = rmm::mr::get_per_device_resource_ref(device_id);
+  return mr.allocate_async(
+    size, rmm::CUDA_ALLOCATION_ALIGNMENT, rmm::cuda_stream_view{static_cast<cudaStream_t>(stream)});
+}
+
+/**
+ * @brief Deallocate memory pointed to by \p ptr.
+ *
+ * @param ptr Pointer to be deallocated
+ * @param size The number of bytes in the allocation
+ * @param device The device whose memory resource one should use
+ * @param stream CUDA stream to perform deallocation on
+ */
+extern "C" void deallocate(void* ptr, std::size_t size, int device, void* stream)
+{
+  rmm::cuda_device_id const device_id{device};
+  rmm::cuda_set_device_raii with_device{device_id};
+  auto mr = rmm::mr::get_per_device_resource_ref(device_id);
+  mr.deallocate_async(ptr,
+                      size,
+                      rmm::CUDA_ALLOCATION_ALIGNMENT,
+                      rmm::cuda_stream_view{static_cast<cudaStream_t>(stream)});
+}

python/rmm/rmm/cpp/cpp_cuda_stream.pxd

@@ -0,0 +1,28 @@
+# 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from cuda.ccudart cimport cudaStream_t
+from libcpp cimport bool
+
+from rmm.cpp.cpp_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()

python/rmm/rmm/cpp/cpp_cuda_stream_pool.pxd

@@ -0,0 +1,23 @@
+# 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from rmm.cpp.cpp_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()

python/rmm/rmm/cpp/cpp_cuda_stream_view.pxd

@@ -0,0 +1,32 @@
+# Copyright (c) 2020, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# 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

python/rmm/rmm/cpp/cpp_device_buffer.pxd

@@ -0,0 +1,58 @@
+# Copyright (c) 2019-2024, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from rmm.cpp.cpp_cuda_stream_view cimport cuda_stream_view
+from rmm.cpp.cpp_memory_resource cimport device_memory_resource
+
+
+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()

python/rmm/rmm/cpp/cpp_device_uvector.pxd

@@ -0,0 +1,39 @@
+# 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from rmm.cpp.cuda_stream_view cimport cuda_stream_view
+from rmm.cpp.device_buffer cimport device_buffer
+from rmm.cpp.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()

python/rmm/rmm/cpp/cpp_logger.pyx

@@ -0,0 +1,66 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from libcpp cimport bool
+
+
+cdef extern from "spdlog/common.h" namespace "spdlog::level" nogil:
+    cpdef enum logging_level "spdlog::level::level_enum":
+        """
+        The debug logging level for RMM.
+
+        Debug logging prints messages to a log file. See
+        `Debug Logging <https://github.com/rapidsai/rmm#debug-logging>`_
+        for more information.
+
+        Valid levels, in decreasing order of verbosity, are TRACE, DEBUG,
+        INFO, WARN, ERR, CRITICAL, and OFF. Default is INFO.
+
+        Examples
+        --------
+        >>> import rmm
+        >>> rmm.logging_level.DEBUG
+        <logging_level.DEBUG: 1>
+        >>> rmm.logging_level.DEBUG.value
+        1
+        >>> rmm.logging_level.DEBUG.name
+        'DEBUG'
+
+        See Also
+        --------
+        set_logging_level : Set the debug logging level
+        get_logging_level : Get the current debug logging level
+        """
+        TRACE "spdlog::level::trace"
+        DEBUG "spdlog::level::debug"
+        INFO "spdlog::level::info"
+        WARN "spdlog::level::warn"
+        ERR "spdlog::level::err"
+        CRITICAL "spdlog::level::critical"
+        OFF "spdlog::level::off"
+
+
+cdef extern from "spdlog/spdlog.h" namespace "spdlog" nogil:
+    cdef cppclass spdlog_logger "spdlog::logger":
+        spdlog_logger() except +
+        void set_level(logging_level level)
+        logging_level level()
+        void flush() except +
+        void flush_on(logging_level level)
+        logging_level flush_level()
+        bool should_log(logging_level msg_level)
+
+
+cdef extern from "rmm/logger.hpp" namespace "rmm" nogil:
+    cdef spdlog_logger& logger() except +