CHANGELOG.md

RMM 0.19.0 (Date TBD)

Please see https://github.com/rapidsai/rmm/releases/tag/v0.19.0a for the latest changes to this development branch.

RMM 0.18.0 (24 Feb 2021)

Breaking Changes 🚨

• Remove DeviceBuffer synchronization on default stream (#650) @pentschev
• Add a Stream class that wraps CuPy/Numba/CudaStream (#636) @shwina

Bug Fixes 🐛

• SetGPUArchs updated to work around a CMake FindCUDAToolkit issue (#695) @robertmaynard
• Remove duplicate conda build command (#670) @raydouglass
• Update CMakeLists.txt VERSION to 0.18.0 (#665) @trxcllnt
• Fix wrong attribute names leading to DEBUG log build issues (#653) @pentschev

Documentation 📖

• Correct inconsistencies in README and CONTRIBUTING docs (#682) @robertmaynard
• Enable tag generation for doxygen (#672) @ajschmidt8
• Document that managed_memory_resource does not work with NVIDIA vGPU (#656) @harrism

New Features 🚀

• Enabling/disabling logging after initialization (#678) @shwina
• cuda_async_memory_resource built on cudaMallocAsync (#676) @harrism
• Create labeler.yml (#669) @jolorunyomi
• Expose the version string in C++ and Python (#666) @hcho3
• Add a CUDA stream pool (#659) @harrism
• Add a Stream class that wraps CuPy/Numba/CudaStream (#636) @shwina

Improvements 🛠️

• Update stale GHA with exemptions & new labels (#707) @mike-wendt
• Add GHA to mark issues/prs as stale/rotten (#700) @Ethyling
• Auto-label PRs based on their content (#691) @ajschmidt8
• Prepare Changelog for Automation (#688) @ajschmidt8
• Build.sh use cmake --build to drive build system invocation (#686) @robertmaynard
• Fix failed automerge (#683) @harrism
• Auto-label PRs based on their content (#681) @jolorunyomi
• Build RMM tests/benchmarks with -Wall flag (#674) @trxcllnt
• Remove DeviceBuffer synchronization on default stream (#650) @pentschev
• Simplify rmm::exec_policy and refactor Thrust support (#647) @harrism

RMM 0.17.0 (10 Dec 2020)

New Features

• PR #609 Adds polymorphic_allocator and stream_allocator_adaptor
• PR #596 Add tracking_memory_resource_adaptor to help catch memory leaks
• PR #608 Add stream wrapper type
• PR #632 Add RMM Python docs

Improvements

• PR #604 CMake target cleanup, formatting, linting
• PR #599 Make the arena memory resource work better with the producer/consumer mode
• PR #612 Drop old Python device_array* API
• PR #603 Always test both legacy and per-thread default stream
• PR #611 Add a note to the contribution guide about requiring 2 C++ reviewers
• PR #615 Improve gpuCI Scripts
• PR #627 Cleanup gpuCI Scripts
• PR #635 Add Python docs build to gpuCI

Bug Fixes

• PR #592 Add auto_flush to make_logging_adaptor
• PR #602 Fix device_scalar and its tests so that they use the correct CUDA stream
• PR #621 Make rmm::cuda_stream_default a constexpr
• PR #625 Use librmm conda artifact when building rmm conda package
• PR #631 Force local conda artifact install
• PR #634 Fix conda uploads
• PR #639 Fix release script version updater based on CMake reformatting
• PR #641 Fix adding "LANGUAGES" after version number in CMake in release script

RMM 0.16.0 (21 Oct 2020)

New Features

• PR #529 Add debug logging and fix multithreaded replay benchmark
• PR #560 Remove deprecated get/set_default_resource APIs
• PR #543 Add an arena-based memory resource
• PR #580 Install CMake config with RMM
• PR #591 Allow the replay bench to simulate different GPU memory sizes
• PR #594 Adding limiting memory resource adaptor

Improvements

• PR #474 Use CMake find_package(CUDAToolkit)
• PR #477 Just use None for strides in DeviceBuffer
• PR #528 Add maximum_pool_size parameter to reinitialize API
• PR #532 Merge free lists in pool_memory_resource to defragment before growing from upstream
• PR #537 Add CMake option to disable deprecation warnings
• PR #541 Refine CMakeLists.txt to make it easy to import by external projects
• PR #538 Upgrade CUB and Thrust to the latest commits
• PR #542 Pin conda spdlog versions to 1.7.0
• PR #550 Remove CXX11 ABI handling from CMake
• PR #578 Switch thrust to use the NVIDIA/thrust repo
• PR #553 CMake cleanup
• PR #556 By default, don't create a debug log file unless there are warnings/errors
• PR #561 Remove CNMeM and make RMM header-only
• PR #565 CMake: Simplify gtest/gbench handling
• PR #566 CMake: use CPM for thirdparty dependencies
• PR #568 Upgrade googletest to v1.10.0
• PR #572 CMake: prefer locally installed thirdparty packages
• PR #579 CMake: handle thrust via target
• PR #581 Improve logging documentation
• PR #585 Update ci/local/README.md
• PR #587 Replaced move with std::move
• PR #588 Use installed C++ RMM in python build
• PR #601 Make maximum pool size truly optional (grow until failure)

Bug Fixes

• PR #545 Fix build to support using clang as the host compiler
• PR #534 Fix pool_memory_resource failure when init and max pool sizes are equal
• PR #546 Remove CUDA driver linking and correct NVTX macro.
• PR #569 Correct device_scalar::set_value to pass host value by reference to avoid copying from invalid value
• PR #559 Fix align_down to only change unaligned values.
• PR #577 Fix CMake LOGGING_LEVEL issue which caused verbose logging / performance regression.
• PR #582 Fix handling of per-thread default stream when not compiled for PTDS
• PR #590 Add missing CODE_OF_CONDUCT.md
• PR #595 Fix pool_mr example in README.md

RMM 0.15.0 (26 Aug 2020)

New Features

• PR #375 Support out-of-band buffers in Python pickling
• PR #391 Add get_default_resource_type
• PR #396 Remove deprecated RMM APIs
• PR #425 Add CUDA per-thread default stream support and thread safety to pool_memory_resource
• PR #436 Always build and test with per-thread default stream enabled in the GPU CI build
• PR #444 Add owning_wrapper to simplify lifetime management of resources and their upstreams
• PR #449 Stream-ordered suballocator base class and per-thread default stream support and thread safety for fixed_size_memory_resource
• PR #450 Add support for new build process (Project Flash)
• PR #457 New binning_memory_resource (replaces hybrid_memory_resource and fixed_multisize_memory_resource).
• PR #458 Add get/set_per_device_resource to better support multi-GPU per process applications
• PR #466 Deprecate CNMeM.
• PR #489 Move cudf._cuda into rmm._cuda
• PR #504 Generate gpu.pxd based on cuda version as a preprocessor step
• PR #506 Upload rmm package per version python-cuda combo

Improvements

• PR #428 Add the option to automatically flush memory allocate/free logs
• PR #378 Use CMake FetchContent to obtain latest release of cub and thrust
• PR #377 A better way to fetch spdlog
• PR #372 Use CMake FetchContent to obtain cnmem instead of git submodule
• PR #382 Rely on NumPy arrays for out-of-band pickling
• PR #386 Add short commit to conda package name
• PR #401 Update get_ipc_handle() to use cuda driver API
• PR #404 Make all memory resources thread safe in Python
• PR #402 Install dependencies via rapids-build-env
• PR #405 Move doc customization scripts to Jenkins
• PR #427 Add DeviceBuffer.release() cdef method
• PR #414 Add element-wise access for device_uvector
• PR #421 Capture thread id in logging and improve logger testing
• PR #426 Added multi-threaded support to replay benchmark
• PR #429 Fix debug build and add new CUDA assert utility
• PR #435 Update conda upload versions for new supported CUDA/Python
• PR #437 Test with pickle5 (for older Python versions)
• PR #443 Remove thread safe adaptor from PoolMemoryResource
• PR #445 Make all resource operators/ctors explicit
• PR #447 Update Python README with info about DeviceBuffer/MemoryResource and external libraries
• PR #456 Minor cleanup: always use rmm/-prefixed includes
• PR #461 cmake improvements to be more target-based
• PR #468 update past release dates in changelog
• PR #486 Document relationship between active CUDA devices and resources
• PR #493 Rely on C++ lazy Memory Resource initialization behavior instead of initializing in Python

Bug Fixes

• PR #433 Fix python imports
• PR #400 Fix segfault in RANDOM_ALLOCATIONS_BENCH
• PR #383 Explicitly require NumPy
• PR #398 Fix missing head flag in merge_blocks (pool_memory_resource) and improve block class
• PR #403 Mark Cython memory_resource_wrappers extern as nogil
• PR #406 Sets Google Benchmark to a fixed version, v1.5.1.
• PR #434 Fix issue with incorrect docker image being used in local build script
• PR #463 Revert cmake change for cnmem header not being added to source directory
• PR #464 More completely revert cnmem.h cmake changes
• PR #473 Fix initialization logic in pool_memory_resource
• PR #479 Fix usage of block printing in pool_memory_resource
• PR #490 Allow importing RMM without initializing CUDA driver
• PR #484 Fix device_uvector copy constructor compilation error and add test
• PR #498 Max pool growth less greedy
• PR #500 Use tempfile rather than hardcoded path in test_rmm_csv_log
• PR #511 Specify --basetemp for py.test run
• PR #509 Fix missing : before LINE in throw string of RMM_CUDA_TRY
• PR #510 Fix segfault in pool_memory_resource when a CUDA stream is destroyed
• PR #525 Patch Thrust to workaround CUDA_CUB_RET_IF_FAIL macro clearing CUDA errors

RMM 0.14.0 (03 Jun 2020)

New Features

• PR #317 Provide External Memory Management Plugin for Numba
• PR #362 Add spdlog as a dependency in the conda package
• PR #360 Support logging to stdout/stderr
• PR #341 Enable logging
• PR #343 Add in option to statically link against cudart
• PR #364 Added new uninitialized device vector type, device_uvector

Improvements

• PR #369 Use CMake FetchContent to obtain spdlog instead of vendoring
• PR #366 Remove installation of extra test dependencies
• PR #354 Add CMake option for per-thread default stream
• PR #350 Add .clang-format file & format all files
• PR #358 Fix typo in rmm_cupy_allocator docstring
• PR #357 Add Docker 19 support to local gpuci build
• PR #365 Make .clang-format consistent with cuGRAPH and cuDF
• PR #371 Add docs build script to repository
• PR #363 Expose memory_resources in Python

Bug Fixes

• PR #373 Fix build.sh
• PR #346 Add clearer exception message when RMM_LOG_FILE is unset
• PR #347 Mark rmmFinalizeWrapper nogil
• PR #348 Fix unintentional use of pool-managed resource.
• PR #367 Fix flake8 issues
• PR #368 Fix clang-format missing comma bug
• PR #370 Fix stream and mr use in device_buffer methods
• PR #379 Remove deprecated calls from synchronization.cpp
• PR #381 Remove test_benchmark.cpp from cmakelists
• PR #392 SPDLOG matches other header-only acquisition patterns

RMM 0.13.0 (31 Mar 2020)

New Features

• PR #253 Add frombytes to convert bytes-like to DeviceBuffer
• PR #252 Add __sizeof__ method to DeviceBuffer
• PR #258 Define pickling behavior for DeviceBuffer
• PR #261 Add __bytes__ method to DeviceBuffer
• PR #262 Moved device memory resource files to mr/device directory
• PR #266 Drop rmm.auto_device
• PR #268 Add Cython/Python copy_to_host and to_device
• PR #272 Add host_memory_resource.
• PR #273 Moved device memory resource tests to device/ directory.
• PR #274 Add copy_from_host method to DeviceBuffer
• PR #275 Add copy_from_device method to DeviceBuffer
• PR #283 Add random allocation benchmark.
• PR #287 Enabled CUDA CXX11 for unit tests.
• PR #292 Revamped RMM exceptions.
• PR #297 Use spdlog to implement logging_resource_adaptor.
• PR #303 Added replay benchmark.
• PR #319 Add thread_safe_resource_adaptor class.
• PR #314 New suballocator memory_resources.
• PR #330 Fixed incorrect name of stream_free_blocks_ debug symbol.
• PR #331 Move to C++14 and deprecate legacy APIs.

Improvements

• PR #246 Type DeviceBuffer arguments to __cinit__
• PR #249 Use DeviceBuffer in device_array
• PR #255 Add standard header to all Cython files
• PR #256 Cast through uintptr_t to cudaStream_t
• PR #254 Use const void* in DeviceBuffer.__cinit__
• PR #257 Mark Cython-exposed C++ functions that raise
• PR #269 Doc sync behavior in copy_ptr_to_host
• PR #278 Allocate a bytes object to fill up with RMM log data
• PR #280 Drop allocation/deallocation of offset
• PR #282 DeviceBuffer use default constructor for size=0
• PR #296 Use CuPy's UnownedMemory for RMM-backed allocations
• PR #310 Improve device_buffer allocation logic.
• PR #309 Sync default stream in DeviceBuffer constructor
• PR #326 Sync only on copy construction
• PR #308 Fix typo in README
• PR #334 Replace rmm_allocator for Thrust allocations
• PR #345 Remove stream synchronization from device_scalar constructor and set_value

Bug Fixes

• PR #298 Remove RMM_CUDA_TRY from cuda_event_timer destructor
• PR #299 Fix assert condition blocking debug builds
• PR #300 Fix host mr_tests compile error
• PR #312 Fix libcudf compilation errors due to explicit defaulted device_buffer constructor

RMM 0.12.0 (04 Feb 2020)

New Features

• PR #218 Add _DevicePointer
• PR #219 Add method to copy device_buffer back to host memory
• PR #222 Expose free and total memory in Python interface
• PR #235 Allow construction of DeviceBuffer with a stream

Improvements

• PR #214 Add codeowners
• PR #226 Add some tests of the Python DeviceBuffer
• PR #233 Reuse the same CUDA_HOME logic from cuDF
• PR #234 Add missing size_t in DeviceBuffer
• PR #239 Cleanup DeviceBuffer's __cinit__
• PR #242 Special case 0-size DeviceBuffer in tobytes
• PR #244 Explicitly force DeviceBuffer.size to an int
• PR #247 Simplify casting in tobytes and other cleanup

Bug Fixes

• PR #215 Catch polymorphic exceptions by reference instead of by value
• PR #221 Fix segfault calling rmmGetInfo when uninitialized
• PR #225 Avoid invoking Python operations in c_free
• PR #230 Fix duplicate symbol issues with copy_to_host
• PR #232 Move copy_to_host doc back to header file

RMM 0.11.0 (11 Dec 2019)

New Features

• PR #106 Added multi-GPU initialization
• PR #167 Added value setter to device_scalar
• PR #163 Add Cython bindings to device_buffer
• PR #177 Add __cuda_array_interface__ to DeviceBuffer
• PR #198 Add rmm.rmm_cupy_allocator()

Improvements

• PR #161 Use std::atexit to finalize RMM after Python interpreter shutdown
• PR #165 Align memory resource allocation sizes to 8-byte
• PR #171 Change public API of RMM to only expose reinitialize(...)
• PR #175 Drop cython from run requirements
• PR #169 Explicit stream argument for device_buffer methods
• PR #186 Add nbytes and len to DeviceBuffer
• PR #188 Require kwargs in DeviceBuffer's constructor
• PR #194 Drop unused imports from device_buffer.pyx
• PR #196 Remove unused CUDA conda labels
• PR #200 Simplify DeviceBuffer methods

Bug Fixes

• PR #174 Make device_buffer default ctor explicit to work around type_dispatcher issue in libcudf.
• PR #170 Always build librmm and rmm, but conditionally upload based on CUDA / Python version
• PR #182 Prefix DeviceBuffer's C functions
• PR #189 Drop __reduce__ from DeviceBuffer
• PR #193 Remove thrown exception from rmm_allocator::deallocate
• PR #224 Slice the CSV log before converting to bytes

RMM 0.10.0 (16 Oct 2019)

New Features

• PR #99 Added device_buffer class
• PR #133 Added device_scalar class

Improvements

• PR #123 Remove driver install from ci scripts
• PR #131 Use YYMMDD tag in nightly build
• PR #137 Replace CFFI python bindings with Cython
• PR #127 Use Memory Resource classes for allocations

Bug Fixes

• PR #107 Fix local build generated file ownerships
• PR #110 Fix Skip Test Functionality
• PR #125 Fixed order of private variables in LogIt
• PR #139 Expose _make_finalizer python API needed by cuDF
• PR #142 Fix ignored exceptions in Cython
• PR #146 Fix rmmFinalize() not freeing memory pools
• PR #149 Force finalization of RMM objects before RMM is finalized (Python)
• PR #154 Set ptr to 0 on rmm::alloc error
• PR #157 Check if initialized before freeing for Numba finalizer and use weakref instead of atexit

RMM 0.9.0 (21 Aug 2019)

New Features

• PR #96 Added device_memory_resource for beginning of overhaul of RMM design
• PR #103 Add and use unified build script

Improvements

• PR #111 Streamline CUDA_REL environment variable
• PR #113 Handle ucp.BufferRegion objects in auto_device

Bug Fixes

...

RMM 0.8.0 (27 June 2019)

New Features

• PR #95 Add skip test functionality to build.sh

Improvements

...

Bug Fixes

• PR #92 Update docs version

RMM 0.7.0 (10 May 2019)

New Features

• PR #67 Add random_allocate microbenchmark in tests/performance
• PR #70 Create conda environments and conda recipes
• PR #77 Add local build script to mimic gpuCI
• PR #80 Add build script for docs

Improvements

• PR #76 Add cudatoolkit conda dependency
• PR #84 Use latest release version in update-version CI script
• PR #90 Avoid using c++14 auto return type for thrust_rmm_allocator.h

Bug Fixes

• PR #68 Fix signed/unsigned mismatch in random_allocate benchmark
• PR #74 Fix rmm conda recipe librmm version pinning
• PR #72 Remove unnecessary _BSD_SOURCE define in random_allocate.cpp

RMM 0.6.0 (18 Mar 2019)

New Features

• PR #43 Add gpuCI build & test scripts
• PR #44 Added API to query whether RMM is initialized and with what options.
• PR #60 Default to CXX11_ABI=ON

Improvements

Bug Fixes

• PR #58 Eliminate unreliable check for change in available memory in test
• PR #49 Fix pep8 style errors detected by flake8

RMM 0.5.0 (28 Jan 2019)

New Features

• PR #2 Added CUDA Managed Memory allocation mode

Improvements

• PR #12 Enable building RMM as a submodule
• PR #13 CMake: Added CXX11ABI option and removed Travis references
• PR #16 CMake: Added PARALLEL_LEVEL environment variable handling for GTest build parallelism (matches cuDF)
• PR #17 Update README with v0.5 changes including Managed Memory support

Bug Fixes

• PR #10 Change cnmem submodule URL to use https
• PR #15 Temporarily disable hanging AllocateTB test for managed memory
• PR #28 Fix invalid reference to local stack variable in rmm::exec_policy

RMM 0.4.0 (20 Dec 2018)

New Features

• PR #1 Spun off RMM from cuDF into its own repository.

Improvements

• CUDF PR #472 RMM: Created centralized rmm::device_vector alias and rmm::exec_policy
• CUDF PR #465 Added templated C++ API for RMM to avoid explicit cast to void**

Bug Fixes

RMM was initially implemented as part of cuDF, so we include the relevant changelog history below.

cuDF 0.3.0 (23 Nov 2018)

New Features

• PR #336 CSV Reader string support

Improvements

• CUDF PR #333 Add Rapids Memory Manager documentation
• CUDF PR #321 Rapids Memory Manager adds file/line location logging and convenience macros

Bug Fixes

cuDF 0.2.0 and cuDF 0.1.0

These were initial releases of cuDF based on previously separate pyGDF and libGDF libraries. RMM was initially implemented as part of libGDF.