This is the release note of v8.0.0b4. See here for the complete list of solved issues and merged PRs.

Highlights

CuPy v8.0.0b4 focuses on performance improvements by adding a general CUB based reduction kernel contributed by @leofang (#3244). We also introduce support for the upcoming CUDA 11 (#3405) although we don’t provide wheels for it yet. Last but not least, several new routines are added to improve the NumPy and SciPy functions coverage.

Changes without compatibility

Change the behavior of dia_matrix.diagonal to follow SciPy 1.5.0 specification. It does not raise ValueError for invalid values anymore. Now an empty array is returned instead. (#3469)

New Features

• Add cupy.shape (#3229)
• CUB-backed _SimpleReductionKernel (#3244, thanks @leofang!)
• Add cupyx.scipy.ndimage sum, mean, standard deviation and variance (#3259, thanks @niteya-shah!)
• Support C++ template code in cupy.RawModule (#3319, thanks @leofang!)
• Add cupy.piecewise (#3329, thanks @Dahlia-Chehata!)
• Add cupy.trim_zeros (#3340, thanks @Dahlia-Chehata!)
• Add cupy.sort_complex (#3348, thanks @Dahlia-Chehata!)
• Add cupy.who (#3361)
• Support cudaDeviceGetLimit / cudaDeviceSetLimit (#3387, thanks @leofang!)
• Add polycompanion (#3398, thanks @Dahlia-Chehata!)
• Add wrappers for cusolverDn<t>potrfBatched and cusolverDn<t>potrsBatched (#3399, thanks @IvanYashchuk!)
• Add polyvander (#3404, thanks @Dahlia-Chehata!)
• Support CUDA 11.0 (#3405)
• Add cupy.shares_memory (#3432)
• Detect and show Thrust build version (#3444, thanks @leofang!)

Enhancements

• Refactor cuTENSOR handle initialization (#2772)
• Deprecate testing.numpy_cupy_raises (#3098)
• Align vector access with #3020 #3022 (#3228)
• Get arch per device and support CUDA 9.2+ (#3366, thanks @leofang!)
• Fix cuTENSOR routines to raise ValueError for invalid arguments (#3374)
• Support ignore_error in kernel optimization (#3410)
• Support boolean in cupyx.scipy.ndimage stats functions (#3419)
• Raise TypeError in cupy.ndarray.__array__ (#3421)
• Make Optuna optional to allow import (#3427)
• Implement flatiter.copy() (#3442)

Performance Improvements

• Speed up CSR SpMV by orders of magnitude (#3430, thanks @leofang!)
• Index CArray using 32-bit indexes (#3448)

Bug Fixes

• Assert that all the pointers are in the same device in concatenate (#3285)
• Fix _count_non_nan datatype for windows (#3350)
• Fix cupyx.time.repeat to accumulate duration after GPU synchronization (#3375)
• Fix PerfCaseResult changing _ts (#3400)
• Fix intermediate dtypes for float16 inputs in cupyx.scipy.ndimage stats functions (#3402)
• Properly reset current stream in case null stream is destroyed (#3423)
• Fix cupy.power(0j, 0j) (#3449)
• Fix TypeError in parameterize test catching CUDADriverError (#3451)
• Fix scipy.dia_matrix.diagonal for scipy==1.5.0 (#3469)

Code Fixes

• Fix array() for readability (#2935)
• Remove unnecessary comparison in cupy.linalg.svd (#3373)
• Fix initial values in cupy._environment (#3413, thanks @leofang!)
• Use find_packages in setup.py (#3424)
• Refactor CUB-backed _SimpleReductionKernel (#3443)

Documentation

• Add documentation for cupyx.optimizing.optimize (#3397)
• Fix sphinx version for travis (#3416)
• Document cupy.fromfile (#3439, thanks @jakirkham!)
• Fix typos in cupy.linalg.det docstring (#3456, thanks @grlee77!)
• Fix docstring of tofile() (#3460, thanks @leofang!)

Installation

• Add optuna and remove theano for doctest requirement (#3446)

Tests

• Add tests for cupy.cuda.cub (#2598, thanks @leofang!)
• Remove chainercv CI configs (#3055)
• Add a test to cover accepting large-size arrays via __cuda_array_interface__ (#3297, thanks @leofang!)
• Add __init__.py to allow importing test packages (#3395)
• Fix ChainerCV tests failing in master branch (#3411)
• Test CUB SpMV (#3428, thanks @leofang!)
• Deprecate testing.empty (#3438)
• Skip some RawModule tests for wrong condition (#3453)
• Use unittest.mock (#3468)

Others

• Bump version to v8.0.0b4 (#3481)