Cupy

This is the release note of v8.4.0. See [here](https://github.com/cupy/cupy/milestone/90?closed=1) for the complete list of solved issues and merged PRs.

Highlights

Gitter Community

We are running a [Gitter chat](https://gitter.im/cupy/community) for general discussions and quick questions. Feel free to join the channel to talk with developers and users!

Changes without compatibility

Removal of older pre-release packages from PyPI

As announced in 4360, we have removed pre-release wheels earlier than v6.0.0rc1 from PyPI. Those version wheels can be found at the [GitHub release](https://github.com/cupy/cupy/releases) page of every version, and can be installed by specifying `-f` option:


pip install --pre cupy-cuda101 -f https://github.com/cupy/cupy/releases/v6.0.0rc1


Changes


Enhancements


- Import DLPack header file & Fix multiple issues (4535)
- Fix sparse format of `kron` (4547)
- Fix return type of `polynomial.__eq__` (4555)

Bug Fixes

- Fix dev info allocation (4501)
- Use `--device-c` for RDC compile (4505)
- Fix `cupy.concatenate` typecheck for out with different dtype (4528)
- Fix `cupy.take` from an empty array (4542)
- Fix integer GEMM (4551)

Tests

- Test `FutureWarning` (4510)

Contributors

The CuPy Team would like to thank all those who contributed to this release!

leofang mor2code

This is the release note of v8.3.0. See [here](https://github.com/cupy/cupy/milestone/87?closed=1) for the complete list of solved issues and merged PRs.


Changes


Enhancements

- Inherit environment variable and detect cl.exe automatically (4417)
- Update CUDA Array Interface to v3 - Part 1 (4446)

Bug Fixes

- Fix `cupy.random.bytes` not working (4323)
- Fix `rcond` arg of `linalg.lstsq` (4408)
- Fix `linalg.lstsq` for complex types (4426)
- Fix `cupy.searchsorted` on HIP (4447)
- Fix out-of-bound access in ndimage rank filters (4449)
- Support complex types in `solve_triangular` (4459)

Code Fixes

- Rename submodules under `cupy.lib` (4353)
- Make names of test classes start with `Test` (4372)

Documentation

- Update links to forums in README (4346)
- Fix comment in docs/source/reference/statistics.rst (4386)
- add `scipy.fft` module to the API comparison table (4391)
- Fix docs of `cupy.random` functions/methods (4474)

Installation

- Fix parallel build (4349)
- Reset `extra_compile_args` for each module (4384)
- Disentangle HIP from CUDA in the build script (4430)
- Add support for cuTENSOR 1.2.2 (4462)

Tests

- Remove travis (4376)
- Refactor test of `linalg.lstsq` (4425)
- Update `[jenkins]` requirement (4473)
- Exclude unsupported dtypes for `TestOrderFilter` (4480)

Others

- Configure Mergify to check GitHub Actions instead of Travis (4381)
- Bump version to `v8.3.0` (4500)

Contributors

The CuPy Team would like to thank all those who contributed to this release!

anaruse grlee77 leofang

This is the release note of v8.2.0. See [here](https://github.com/cupy/cupy/milestone/85?closed=1) for the complete list of solved issues and merged PRs.

Changes

Enhancements

- Record Cython build version (4188)
- Add parallel build feature (4273)
- Bump cuDNN to v8.0.5 (4313)
- Defer import in `cupy/_environment.py` (4329)

Bug Fixes

- Fix broadcasting behavior in `ndimage.measurements functions` (4204)
- Refactor `AssertFunctionIsCalled` (4253)

Code Fixes

- Rename submodules under `cupyx.linalg` package (4202)
- Use `assert` statement instead of `self.assert*` methods (4297)

Documentation

- Add cupy-cuda111 to README (4212)
- Add missing functions to the API reference (4257)
- cupy-cuda111 package now on PyPI (4335)

Tests

- Fix tests of `__bytes__` (4255)
- Fix `numpy_cupy_equal` for case that both numpy cupy raise errors (4260)
- Use GitHub Actions (4286)
- Skip some failing tests for fp16 + CUDA 9.0 (4324)
- Add import test for ROCm (4334)

Others

- Bump version to v8.2.0 (4332)
- ROCm: Support hipCUB/rocPRIM (4327)
- Fix output dtype of `linalg.norm` (4230)
- Warn non-tuple sequence for multidimensional indexing (4285)

Contributors

The CuPy Team would like to thank all those who contributed to this release!

grlee77 leofang

This is the release note of v8.1.0. See [here](https://github.com/cupy/cupy/milestone/83?closed=1) for the complete list of solved issues and merged PRs.

Highlights

CUDA 11.1 Support

Support for CUDA 11.1 is added in 4184, with CUDA 11.1, GeForce RTX 30 series and Quadro RTX series can now be used in CuPy.

Notes on Wheel Packages

Update (2020-11-25): `cupy-cuda111` is now available on PyPI.
~CuPy for CUDA 11.1 (`cupy-cuda111`) wheel packages are currently only available for Windows. We are going to publish Linux wheels once we get [approval](https://github.com/pypa/pypi-support/issues/690) from the PyPI team. Meanwhile, Linux wheels can be downloaded from the Assets section below (or `pip install cupy-cuda111 -f https://github.com/cupy/cupy/releases/tag/v8.1.0`).~

New Features

- Add sparse pointwise equality & inequality functions (4004)
- Add `cudaGetDeviceProperties` (4103)
- Add `order` option in `cupy.testing.shaped_random` (4104)
- Add support for CUDA 11.1 (4191)

Enhancements

- Bump cuDNN to v8.0.4 (4069)
- Show numpy and scipy versions in `show_config` (4079)
- Support pickling `cupy.RawKernel` (4154)

Bug Fixes

- Fix `csr2csc` for zero-size matrix (3922)
- Add a kernel for integer GEMM (4067)
- Fix potential segfault when reduction axis is empty (4068)
- Workaround cudaPointerGetAttributes error in CUDA 10.2+ (4089)
- Add work-around for issue in cutensorReduction of cuTENSOR 1.2.1 (4098)
- Fix `argmax` and `argmin` for F-order inputs (4106)
- Fix CUB block reduction for F-order arrays with ndim > 2 (4109)
- ROCm: Fix `getDeviceProperties` for HIP (4113)
- Fix `argmax`/`argmin` in CUB block reduction for F-order arrays with ndim > 1 (4115)
- Fix typos in `cupy.cuda.cufft` (4117)
- Handle `np.nan` and `np.inf` constant values properly in ndimage functions (4133)
- Fix 64-bit int types in `type_dispatcher.cuh` (4134)
- Add `compute_35` for CUDA 11.0+ (4140)
- Fix device properties for cuda 9.2 (4152)
- fix mode='opencv' case in cupyx.scipy.ndimage.affine_transform (4158)
- Fix `argwhere` for 0d inputs (4174)
- Fix to use current stream properly with CUDA-related libraries (4175)
- Add compute capability checking for cublasGemmEx() (4180)
- Fix cupyx.seterr() when `linalg` not supplied (4189)
- Fix `nonzero` for 0d inputs (4190)

Code Fixes

- Rename submodules under `cupyx.scipy.sparse` (3959)
- Rename submodule under `cupy.fft` package (4066)
- Hide private names in `cupy.cusolver` (4076)
- Move `_normalize_axis_index` to `cupy/core/internal.pyx` (4086)
- Rename `cupyx.rsqrt` submodule (4116)
- Rename submodules under `cupyx.scipy.special` (4119)
- Move `matmul` from `core.pyx` to `_routine_linalg.pyx` (4123)
- Hide private names in `cupy.cutensor` (4147)
- Rename `cupy.manipulation` submodule to `cupy._manipulation` (4181)
- Rename `cupy.io` submodule to `cupy._io` (4183)
- Rename submodule under `cupyx.scipy.fft` (4186)
- Rename submodules under `cupy.linalg` package (4187)

Documentation

- Fix typo (4056)
- Update README and docs for a unified tagline (4074)
- Improve the plan cache documentation (4087)
- Simplify ROCm install guide (4128)

Installation

- Add `CUDA_VERSION` define for Cython compilation (4035)

Tests

- Require SciPy 1.2 for sparse comparison (4041)
- Make parameterized dtype test skip by `pytest.skip` (4179)
- Code fix on tests for `cupyx.scipy.ndiamge` stats functions (4182)
- Fix tests that have side effects (4185)

HIP/ROCm

- ROCm: Fix bugs and test suites to make ROCm/HIP happy - Part 2 (4063)
- ROCm: Build on the latest ROCm (4126)

Others

- Bump version to v8.1.0 (4195)

Contributors

The CuPy Team would like to thank all those who contributed to this release!

anaruse garanews grlee77 leofang

Highlights

The CuPy v8.0.0 release includes a number of new features, as well as enhanced NumPy/SciPy functionality coverage.

* **TensorFloat-32 (TF32) Support**
* CuPy now supports [TensorFloat-32](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/), a new feature available in NVIDIA Ampere GPU and CUDA 11. Set `CUPY_TF32=1` environment variable to boost the performance of matrix multiplications in routines such as `cupy.matmul` or `cupy.tensordot`.

* **Official support for NVIDIA cuTENSOR and CUB libraries**
* Several routines in CuPy now support using the [cuTENSOR](https://docs.nvidia.com/cuda/cutensor/index.html) and [CUB](https://nvlabs.github.io/cub/) libraries to further improve performance. Set `CUPY_ACCELERATORS=cub,cutensor` environment variable to benefit from these libraries.
* **Enhanced kernel fusion**
* While combining multiple kernels into a single one using `cupy.fuse`, it was only possible to use a single reduction operation (`cupy.sum`, etc.) at the end. With the new kernel fusion mechanism available in CuPy v8, now it is possible to combine multiple element-wise operations with interleaved reductions.
* **Automatic tuning of kernel launch parameters**
* CuPy now supports discovering the optimal CUDA kernel launch parameters depending on the data and device properties for better performance. See the API reference ([`cupyx.optimizing.optimize`](https://docs.cupy.dev/en/latest/reference/generated/cupyx.optimizing.optimize.html)) for details.
* **Memory pool sharing with external libraries**
* With the new `PythonFunctionAllocator` API, you can let CuPy use arbitrary Python functions instead of a built-in memory pool when managing GPU memory. This improves interoperability with external libraries; for example, you can flexibly use CuPy to preprocess data or use its custom CUDA kernel features inside PyTorch. With [pytorch-pfn-extras](https://github.com/pfnet/pytorch-pfn-extras) bundled allocator it is possible to [easily use the PyTorch memory pool from CuPy](https://github.com/pfnet/pytorch-pfn-extras/blob/master/docs/cuda.md).
* **Improved NumPy/SciPy function coverage**
* Many functions added, including the NumPy Polynomials package (results of [Google Summer of Code 2020](https://summerofcode.withgoogle.com/archive/2020/projects/5856911817179136/), thanks Dahlia-Chehata!), the SciPy image processing package, and extended support for the SciPy sparse matrices package.

For the list of all backward-incompatible changes in v8, please refer to the [Upgrade Guide](https://docs.cupy.dev/en/latest/upgrade.html#cupy-v8).

Notes on Wheel Packages

* CuPy for CUDA 10.1 (`cupy-cuda101`), 10.2 (`cupy-cuda102`), and 11.0 (`cupy-cuda110`) packages are built with cuDNN v8 support but without bundled cuDNN shared libraries (see 3724 for the discussion). To use cuDNN features, You need to download cuDNN library using the following command: `python -m cupyx.tools.install_library --library cudnn --cuda X.X`. It is also possible to install cuDNN v8.0.x via the system package manager (e.g., `apt install libcudnn8` or `yum install libcudnn8`) or manually install it and set `LD_LIBRARY_PATH` environment variables.

This is the release note of v8.0.0rc1. See [here](https://github.com/cupy/cupy/milestone/78?closed=1) for the complete list of solved issues and merged PRs.

We are planning to release the final v8.0.0 on October 1st. Please start testing your workload with this release. See the [Upgrade Guide](https://docs.cupy.dev/en/v8.0.0rc1/upgrade.html#cupy-v8) for the list of possible breaking changes.

Highlights

* This release adds support for CUDA 11, NumPy 1.19, and SciPy 1.5.
* Several performance improvements when using cuTENSOR, sparse matrices indexing, matrix multiplication with CUDA 11 using TF32.
* Compatibility with `numpy.poly` is being increased thanks to our GSoC student Dahlia-Chehata!
* Added an interface (3126) to support using external memory allocators such as the PyTorch one (https://github.com/pytorch/pytorch/pull/33860).

Notes on Wheel Packages

* Update on 2020-09-23: `cupy-cuda110` package is now available on PyPI! ~CuPy for CUDA 11.0 (`cupy-cuda110`) wheel packages are currently available only for Windows. We are going to publish Linux wheels once we get [approval](https://github.com/pypa/pypi-support/issues/553) from the PyPI team. (Meanwhile, Linux wheels can be downloaded from the Assets section below (or `pip install cupy-cuda110 -f https://github.com/cupy/cupy/releases/tag/v8.0.0rc1`). Those wheels will be removed once we publish the package on PyPI.)~
* CuPy for CUDA 10.1 (`cupy-cuda101`), 10.2 (`cupy-cuda102`), and 11.0 (`cupy-cuda110`) packages are built with cuDNN v8 support but without bundled cuDNN shared libraries (see 3724 for the discussion). To use cuDNN features, You need to download cuDNN library using the following command: `python -m cupyx.tools.install_library --library cudnn --cuda X.X`.
It is also possible to install cuDNN v8.0.x via the system package manager (e.g., `apt install libcudnn8` or `yum install libcudnn8`) or manually install it and set `LD_LIBRARY_PATH` environment variables.


Changes without compatibility

Deprecate `cupy.sparse` package (3839, 3856)
CuPy's sparse matrix support was initially implemented in the `cupy.sparse` package. It was moved to the `cupyx.scipy.sparse` namespace in CuPy v5, while keeping the `cupy.sparse` one for backward compatibility.
Since there is no equivalent package in NumPy, it was decided that it will be deprecated and
eventually removed.

Deprecate `*_enabled` flags under `cupy.cuda` (3732)
Before it was possible to use `cupy.cuda.nccl_enabled` or similar to detect whether NCCL, cuTENSOR or other optional CUDA libraries are available to use. Now this pull-request introduced a per-module flag (`cupy.cuda.nccl.available`, `cupy.cuda.cutensor.available`) to obtain the same information.

Bump version in Docker images (3733)

The current base Docker images have been updated from Ubuntu 16.04, CUDA 9.2, and Python 3.5 to Ubuntu 18.04, CUDA 10.2, and Python 3.6.


New Features

- Add `cupy.ndim` (3060)
- Add `PythonFunctionAllocator` (3126)
- Compressed Sparse Inner Indexing (3486)
- Add `cupy.polyadd` (3548)
- Add `cupy.polymul` (3590)
- Add `cupy.polysub` (3593)
- Add most of `scipy.linalg.special_matrices` (3641)
- Add `scipy.signal` functions that are simple wrappers of `ndimage` functions (3645)
- Add `cupyx.scipy.ndimage.fourier_shift`, `fourier_gaussian`, `fourier_uniform` (3654)
- Add 2D Sparse Slicing (3657)
- Add 2D Sparse Slicing + Row Indexing (3658)
- Add 2D Sparse Slicing + Row & Column Indexing (3659)
- Add `cupy.roots` for Hermitian or symmetric matrix (3703)
- Add `cupy.polyval` (3725)
- Support `__cuda_array_interface__` in `cupy.poly1d` (3729)
- Implement library preloading for wheels (3731)
- Add `cupy.poly1d.__pow__` (3734)
- Add `scipy.signal.convolve` and `correlate` functions (3748)
- Add `trimcoef` (3793)

Enhancements

- Avoid disk I/O in compiler (3164)
- Add check for method in Randomstate seed (3282)
- Support negative `axis` in sparse `min`/`max`/`argmin`/`argmax` (3497)
- Mark `nonzero` parameters experimental in sparse `min`/`max` (3583)
- Add a `compile` method for `RawKernel` and `RawModule` (3644)
- Handle `__cuda_array_interface__` in `asnumpy` (3718)
- Use `cublasGemmEx` in `tensordot_core` when CUDA11 (3719)
- Deprecate `*_enabled` flags under `cupy.cuda` (3732)
- Fix handle types to `intptr_t` (3746)
- Support TF32 (3810)
- Deprecate `cupy.sparse` package (3839)
- Add `path` and `readonly` options to `cupyx.optimizing.optimize` (3845)
- Adding a workaround for even-length inputs to `scipy.signal.sepfir2d` (3750)
- Add multi-axis support to `cupy.flip` (3742)


Performance Improvements

- Speed up `cupy.vdot` (3678)
- Improve `cupy.cutensor` (3700)
- More improvement of `cupy.cutensor` (3744)
- Improve 2D sparse row slicing (3782)
- Improve median_filter, rank_filter and percentile_filter (3813)
- Improve CSR matrix `getrow`, `getcol` and some slicing (3851)

Bug Fixes

- Fix `float16` `ndarray` input in `histogram` with CUB (3617)
- Support order argument in `cupy.ones`, `cupy.full` and `cupy.eye` (3655)
- Work around a known CUB SpMV bug (3679)
- Fix broken message format (3691)
- Fix `can_use_device_segmented_reduce()` for incompatible axes (3740)
- Fix circular imports (3743)
- Skip FFT input checks for some CUDA >= 10.1 cases (3763)
- Fix CUDA 11 multi-GPU FFT bug (3775)
- Temporary fixes for cudnn v8 (3790)
- Fix `cupy.correlate` (3801)
- Copy input by default for C2R transform (3848)
- Fix `cupy.sparse.*` deprecation (3856)
- Fix cub not bundled in wheels (3879)
- Fix wheel not loading bundled cuDNN on Windows (3880)
- Add option to include wheel metadata (3881)
- Fix not to use `cupy.cuda.*` from CuPy codebase (3883)

Code Fixes

- Add `cupy_backends/cuda/libs/cutensor.pxd` (3595)
- Refactor `_make_decorator` in helper.py (3697)
- Refactor `cupy.poly1d` tests (3704)
- Remove unnecessary imports in `cupy._sorting` (3706)
- Rename `cupy.binary` submodule to `cupy._binary` (3707)
- Rename `cupy.creation` submodule to `cupy._creation` (3708)
- Rename `cupy.functional` submodule to `cupy._functional` (3710)
- Rename `cupy.indexing` submodule to `cupy._indexing` (3711)
- Remove unnecessary imports of `cupy.linalg` (3714)
- Rename `cupy.misc` submodule to `cupy._misc` (3726)
- Rename `cupy.padding` submodule to `cupy._padding` (3727)
- Rename submodules under `cupy.random` package (3772)
- Refactor logical routines from `core.pyx` (3804)
- Refactor binary-op routines from `core.pyx` (3816)
- Fix typo (3850)
- Resolve circular imports between `cupy` and `cupyx.scipy` (3854)

Documentation

- Correct format of docstrings in creation routines (3752)
- Update docs for v8 (3802)
- Fix a broken document (3807)
- Add `cupy-cuda110` package to README (3817)
- Fix documents to reflect `CUPY_ACCELERATORS` (3818)
- Support Optuna v2 (install docs) (3842)
- Add upgrade guide for v8 (3863)
- Fix broken link in the installation guide (3864)

Installation

- Bump version in Docker images (3733)
- Update `classifiers` in `setup.py` (3814)
- Install SciPy and Optuna to Docker image (3844)

Tests

- Fix wrong test file name (3722)
- Fix test to run without NCCL (3735)
- Avoid mutation of `os.environ` (3749)
- Relax tolerance in `TestArrayElementwiseOp::test_doubly_broadcasted_pow` (3758)
- More on using `unittest.mock` (3791)
- Fix test to run without cuDNN (3846)

Others

- Bump version to v8.0.0rc1 (3882)
- Make nvrtc `getPTX` use `bytes` instead of `unicode` (3237)
- Add hiprtc support (3238)
- Fix build and import errors for ROCm (3786)

Contributors

The CuPy Team would like to thank all those who contributed to this release!

anaruse, cjnolet, coderforlife, Dahlia-Chehata, jakirkham, leofang, niteya-shah, pentschev