Cupy

This is the release note of v6.1.0. See [here](https://github.com/cupy/cupy/milestone/51?closed=1) for the complete list of solved issues and merged PRs.

Notes

- Tensor Core in cuDNN convolution is tentatively disabled for Turing GPUs due to some test failures. The issue is under investigation and hopefully fixed in a future version.

Enhancements

- Improve `atomicAdd` in `histgram` and `sample` (2217)

Bug Fixes

- Revert "Fix usage check of Tensor Core" (2198)
- Fix `array_split` with non-equally dividing sections (2214)
- Avoid division by zero in tensordot, allowing 0-length arrays (2231, thanks pentschev!)
- Make `RandomState.permutation` compatible with `random.permutation` (2253)

Documentation

- Add cupy-cuda101 to README (2203)
- Fix duplicate object description of cupy (2240)

Installation

- Add setup option to copy include files in wheel (2232)

Tests

- Ignore invalid axis type test in NumPy 1.12.x or earlier (2200)

This is the release note of v6.0.0. See [here](https://github.com/cupy/cupy/milestone/49?closed=1) for the complete list of solved issues and merged PRs.

This release note only covers the difference from v6.0.0rc1; for all highlights and changes, please refer to the release notes of the pre-releases:

This is the release note of v6.0.0rc1. See [here](https://github.com/cupy/cupy/milestone/48?closed=1) for the complete list of solved issues and merged PRs.

Highlights

- CUDA 10.1 and cuDNN 7.5 are now supported. CuPy also starts to compile for compute capability 7.5 for Turing GPUs.
- After this release, the master branch is switched to the development of v7 series. v6.0.0 will continue developing at the `v6` branch.

New Features

- New RNN API introduced in cuDNN v7.2 (1609)
- Add `diff` and `unwrap` (1933, thanks a2kiti!)
- Support fusion feature of `copyto` method (1983)
- Add `lu_factor` and `lu_solve` to `cupyx.scipy.linalg` (2051, thanks msakai!)

Enhancements

- More support `__cuda_array_interface__` (2058)
- Fix fp16 issue in batch normalization (2060)
- Keep backward compatibility on `cupy.cudnn.batch_normalization_forward_training` (2072)
- Check if gc module is still available (2116)
- Support CUDA 10.1 + cuDNN 7.5 + Turing (2117)

Performance Improvements

- Do exact type comparison instead of `isinstance` for `numpy.dtype` (2016)
- Improve `_routines_manipulation` (2038)

Bug Fixes

- Fix assigning from complex to float (1911)
- Ensure that sparse matrix shapes are always a tuple of `int` (1943, thanks grlee77!)
- Avoid sharing handles between threads (2053)
- Check array contiguity in `copy` (2075)
- Support dltensor with strides of NULL (2097, thanks crcrpar!)
- Fix assertion error in `_Chunk.split` (2103, thanks liwt31!)

Code Fixes

- Use single quote (2049)
- Use `assert` in `helper.py` instead of `self.assertXXX` (2077)
- Remove unused variable (2105, thanks crcrpar!)
- Reorganize import in `device.pyx` (2121)

Documentation

- Document `__array_function__` (1979)
- Add NCCL v2.4 support to docs (2065)
- Add strides to docstring of ndarray (2096, thanks crcrpar!)

Installation

- Use deep copy in setting up RPATH (2073)

Tests

- Do not fail by warnings when building docs (2080)
- Test `linalg.cholesky` with more stable inputs (2084)

This is the release note of v6.0.0b3. See [here](https://github.com/cupy/cupy/milestone/45?closed=1) for the complete list of solved issues and merged PRs.

New Features

- Implement `cupy.put` and `cupy.place` (1787, thanks grafi-tt!)
- Add `plan` argument to FFT functions in `cupyx.scipy.fftpack` (1942 2033, thanks leofang!)
- Add `out` argument to `ndarray.get` for asynchronous device-to-host copy (1970, thanks jeng1220)
- Add NCCL 2.4 functions (1992)

Enhancements

- Remove experimental warning of `cupy.fuse` (1379)
- Support comparing complex-number arrays in `cupy.allclose` (1947, thanks leofang!)
- Use version interface in NCCL 2.3.4 (1985)
- Enhance error messages of elementwise operation in fusion mode (2007)
- Fix vector handling (2008)
- Support cuDNN FP16 batch normalization (2034)
- Avoid `import numpy` (2040)
- Avoid using pytest attributes during import (2055)
- Support Python scalars in `iscomplexobj` (1991)

Performance Improvements

- Improve performance of reduction on outer axis (2010, thanks grafi-tt!)
- Reduce vector copy in ndarray initialization (2015)
- Improve kind score operation (2017)
- Improve dictionary operation (2018)
- Improve `get_ufunc_kernel` (2019)
- Reduce `get_device_id` call (2021)
- Improve performance of broadcasting in elementwise kernels (2022)
- Improve `_is_fusing` performance (2023)

Bug Fixes

- Fix `cupy.random.randint` fail with size zero (1967)
- Call free_all_blocks to free CUDA memory (1984)
- Define `__dealloc__` instead of `__del__` for cdef-classes to fix memory leak (1995, thanks msakai!)
- Fix NCCL version assignment for NCCL < 2.3.4 (2009)
- Fix `__array_function__` bug (2024)
- Fix usage array function for CuPy modules (2026, thanks pentschev!)
- Fix random generator seed type (2036)

Code Fixes

- Remove duplicate declaration of `cudaMalloc` (2028, thanks grlee77!)
- Fix incorrect `dtype` for `ipiv` buffer in `cupy.linalg.inv` (2043, thanks msakai!)

Documentation

- Add documentation of `cupy.fuse` (1789)
- Fix README image to use URL (1977)
- Add cupy-cuda100 to README (1978)
- Document NumPy 1.16 support (1986)
- Document new features in `cupy.fft` and `cupyx.scipy.fftpack` (2035, thanks leofang!)
- Fix `cupyx.scipy.get_array_module`'s docstring (2050, thanks msakai!)
- Make `plan` argument in 2D/3D FFT functions experimental (2056)

Installation

- Eliminate RUNPATH to use correct cuDNN (1770)
- Fix eliminate-runpath (2014)
- Fix RPATH not set correctly (2064)

Examples

- Improve memcpy example (1999, thanks grafi-tt!)

Tests

- Fix test of `assert_array_list_equal` (1997)
- Fix test with NumPy 1.16.1 (1998)
- Do not "accept error" if bad kwarg is passed to test itself (2001)

This is the release note of v6.0.0b2. See [here](https://github.com/cupy/cupy/milestone/43?closed=1) for the complete list of solved issues and merged PRs.

Highlights

- Compatibility with NumPy 1.16 and support for the [`__array_function__`](http://www.numpy.org/neps/nep-0018-array-function-protocol.html) interface allowing CuPy arrays to be passed to NumPy functions.

New Features

- Implement `__array_function__` interface (1650)
- Implement SciPy-compatible FFT functions (1745)
- Add `PlanNd` class for faster 2D and 3D FFTs (1746, thanks grlee77!)
- Add `order` argument to` empty_like`,` zeros_like`, etc. (1819, thanks grlee77!)
- Add `cupyx.scipy.sparse.diags` (1840, thanks grlee77!)
- Add the `order` kwarg to `cupy.reshape` and the underlying `ndarray` method (1843, thanks grlee77!)
- Support order kwarg in `asarray`, `asanyarray`, `ndarray.get`, `tonumpy` (1845, thanks grlee77!)
- `cumsum`, `cumprod`: support array-like input (1847, thanks grlee77!)
- Add sparse option to meshgrid (1848, thanks grlee77!)
- Add `attributes` property to `Device` (1869, thanks grlee77!)
- Support new APIs in cuDNN v7.4.1 (1884)
- Add `as_strided` (1897, thanks fujiisoup!)
- Add `fnc` and `forc` to `core.flags.Flags` (1898, thanks grlee77!)
- CuPy external memory pool with memory pool and function pointers (1904)
- Allow multi-axis `roll` as in NumPy 1.12+ (1818, thanks grlee77!)

Enhancements

- Add `cupy.util.PerformanceWarning` (1607)
- Support NumPy arrays as seeds in `RandomState` (1689, thanks mrocklin!)
- Better support cuda-gdb (1773)
- Allow `cupy.fuse` taking any parameters when fails to fuse (1817)
- Verbose broadcast failure message (1836)
- Bundle CUDA fp16 headers (1837)
- Refactor array creation for RNN API (1842)
- Support for documenting NCCL interfaces (1857)
- Use `item` instead of `asscalar` to support NumPy 1.16 (1880)
- Support NumPy 1.16 (1881)
- Allow passing `None` as parameters of `Fusion.__call__` (1965)

Performance Improvements

- Improve `cudnn.py` performance (1375)
- Improve `matmul` perfomance (x3 faster) (1547)
- Improve `matmul` (small changes) (1894)
- Improve `Fusion.__call__` (1832)

Bug Fixes

- Fix backward of batch normalization (1822)
- Fix `RecursionError` bug in `conj` method of sparse matrix classes (1846, thanks grlee77!)
- Allow `None` as `cx` argument in cuDNN RNN functions (1862)
- Fix `ndarray` initialization bug (1907)
- Set `size` attribute of externally allocated memory (1926)
- Fix `ndim` not using `cdef` (1935)
- Ensure the input to the CUFFT `Plan1D` class is C contiguous (1944, thanks grlee77!)
- Fix wrong variable name in `cudnn.pyx` (1966)

Documentation

- Document Numba CUDA array conversion (1786)
- Align with the latest NumPy docs (1804)
- Update docs for newly supported libraries (1854)
- Fix bad hyperlink in `cupy.around` documentation (1856)
- Add memory management documentation (1871)
- Add `diags` to documentation (1872)
- Fix stylecheck installation in contribution guide. (1873, thanks crcrpar!)
- Avoid using `_ufunc_wrapper` and `_reduction_wrapper` (1888)
- Rename Code of Conduct filename (1900)

Tests

- Add tests for complex `min` and `max` (1828)
- Fix `randint` high value in test for Windows (1892)
- Fix test to support SciPy 1.12 (1968)

Code Fixes

- Refactor manipulation routines from `core.pyx` (1620)
- Use language level 3 in cythonize (1792)
- Change argument name of `core.create_comparison` (1829)
- Refactor cyclic imports (1905)
- Refactor math and indexing routines from `core.pyx` (1949)
- Clean up `internal.pyx` (`prod`) (1951)
- Fix int types (1952)
- Rename `External*` to `CFunction*` (1958)
- Refactor logic, sorting and statistics routines from `core.pyx` (1959)
- Use LF instead of CR+LF (1909)

Others

- Ignore W503 and W504 (1953)

This is the release note of v6.0.0b1. See [here](https://github.com/cupy/cupy/milestone/41?closed=1) for the complete list of solved issues and merged PRs.

New Features

- Support comparison operators for complex numbers (1636, thanks themightyoarfish!)
- Implement RNN high level API (1659)
- Support array conversion from numba to cupy w/o memcpy (1760)
- Add `ndarray.item` method (1815)

Enhancements

- Check if `ElementwiseKernel` doesn't broadcast mutable `in_params` (1601)
- Support `ndim` checking in Fusion (1713)
- Remove code for CUDA 7.0 / 7.5 and cuDNN v4 (1733)
- Fix `DropoutStates` interface to avoid using cuDNN handle (1734)
- Add cuDNN pooling interface (1735)
- Add cuDNN softmax interface (1736)
- Implement cuDNN batch normalization interface (1737)
- Assume CUDA v8 in cuSPARSE/cuSOLVER check (1780)
- Rename `_make_rnn_workspace` function (1838)
- Make a helper function to get the number of layers in RNN (1839)

Bug Fixes

- Fix invalid memory access during reduction whose output is large to exceed 32-bit (1774)
- Fix incorrect cast in element-wise op for float array and complex scalar (1795)
- Fix error in typecast from fp16 to complex dtype in `cupy.fuse` (1799)

Documentation

- Improve comparison table (1622)
- Add requirements to build documentation on RTD (1751)
- Add policy on compatibility of random sampling (1754)
- Fix description of `CUDA_PATH` environment variable (1775)
- Indicate alias in comparison table (1791)
- Add code of conduct (1793)
- Fix English in `ndarray.nbytes` documentation (1797)
- Add note about `cupy.random.{get_set}_state` (1805)
- Add docs about conversion from/to cupy ndarray and sparse (1806)
- Enable intersphinx to `cupyx` module (1809, thanks grafi-tt!)
- Add 10.0 to supported CUDA versions (1824)
- Add v7.3 and v7.4 to supported cuDNN versions (1823)
- Fix bad source hyperlinks (1835)

Tests

- Fix `PendingDeprecationWarning` when running test against NumPy 1.15 (1755)
- Support skipping tests decorated by condition or parametrize (1757)
- Fix test failure of binomial distribution in Windows (1758)
- Fix userspace kernel test failure in Windows (1759)

Other

- Add issue template (1808)