This is the release note of v8.0.0b1. See here for the complete list of solved issues and merged PRs.

Known packaging issues:

• CuPy build fails when using CUDA 8.0 on Windows (#3076). Due to this issue, cupy-cuda80 wheel packages for Windows are unavailable for this version. Linux or CUDA 9.0+ users are unaffected.

Highlights

CuPy gets faster and more stable towards its v8.0.0 release. This version adds a handful of new routines, adds library wide performance improvements and corrects several bugs.

Changes without compatibility

• Removed cupy.scatter_add, which had been deprecated since CuPy v4. Use cupyx.scatter_add instead.

New Features

• Add get_global() to cupy.RawModule (#2510, thanks @leofang!)
• Support multi-GPU in cupy.cuda.cufft.Plan1d (#2644, thanks @leofang!)
• Add hstack, vstack, and bmat to cupyx.scipy.sparse (#2665, thanks @cjnolet!)
• Add cupy.require (#3083, thanks @niteya-shah!)
• Add cupy.compress (#3103, thanks @Harshan01!)
• Add cupy.ravel_multi_index (#3104, thanks @grlee77!)
• Add cupy.extract (#3109, thanks @Harshan01!)
• Add cupy.bitwise_not as alias to invert (#3120, thanks @Harshan01!)
• Add cupy.argwhere (#3135, thanks @rushabh-v!)
• Add cupy.select (#3138, thanks @niteya-shah!)
• Add cupy.cuda.ExternalStream (#3141)
• Add cupy.array_equal (#3189, thanks @rushabh-v!)

Enhancements

• Add ndarray variants AND inplace support in fallback_mode (#2391, thanks @Piyush-555!)
• Support array-like start/stop and add axis argument to linspace (#2461, thanks @grlee77!)
• Add fp16 support of CUB (#2600, thanks @y1r!)
• Raise errors instead of assertion on array type checks (#2795)
• Drop support for NumPy 1.15 or earlier (#2938)
• Import using_allocator in cupy.cuda (#2951, thanks @jakirkham!)
• Remove __future__ imports (#2995)
• Support CUB prod (#3067, thanks @leofang!)
• Remove deprecated cupy.scatter_add (#3074)
• Update cupy.pad to use cupy.linspace instead of numpy.linspace internally (#3101, thanks @grlee77!)
• Histogram update: support range, weights and density (#3124, thanks @grlee77!)
• Add support for ord = 2, -2, and 'nuc' in cupy.linalg.norm (#3130, thanks @rushabh-v!)
• Use ElementwiseKernel in cupy.fill_diagonal (#3139)
• Allow dia_matrix creation from SciPy equivalent (#3160, thanks @jakirkham!)
• Add labels in the benchmark and add kwargs to repeat (#3172, thanks @rushabh-v!)
• Add out parameter to cupy.concatenate and cupy.stack (#2983)
• Fix reshape to raise ValueError for order 'K' (#3123)

Performance Improvements

• Improve cuDNN performance when using deterministic mode (#1380)
• Improve performance of cumsum and cumprod (#2907)
• Improve ndimage convolve and correlate (#3179)
• Add check of c_contiguous when indexing CArray (#3191)

Bug Fixes

• Fix an issue using non-existing attribute in cub.pyx (#2985)
• Use size_t nbytes in __cuda_array_interface__ (#3009, thanks @jakirkham!)
• Fix fill_diagonal (#3011)
• Fix cupy.random.multivariate_normal (#3018, thanks @espg!)
• Use Python scalar as random seed (#3054)
• Properly decrement total bytes in memory pool (#3068)
• Fix condition to use slice copy in ndarray.__setitem__ (#3088)
• Fix compiler bug when building cupy.cuda.cub with CUDA < 9.2 (#3089, thanks @leofang!)
• Fix cub_reduction for CUPY_CUB_MIN and float16 arrays (#3100)
• Use time.process_time instead of time.clock (#3128, thanks @rushabh-v!)
• Add support for 0 sized matrices in svd (#3140, thanks @rushabh-v!)
• Fix CUB-based cupy.prod for half precision (#3148, thanks @leofang!)
• Fix error type and message in coo_matrix (#3150)
• Allow MatDescriptor to be pickle-able (#3157, thanks @jakirkham!)
• Fix erfinv & erfcinv in cupyx.scipy.special (#3159, thanks @leofang!)
• Remove some xfails in sorting tests (#3167)
• Fix Event.__del__ behavior on shutdown` (#3176)
• Add the missing initialization value in the reduction test (#3194, thanks @leofang!)

Code Fixes

• Clean up internal.pyx (get_contiguous_strides) (#1950)
• Remove custom tempdir context manager (#3003)
• Use intptr_t instead of size_t for cuSPARSE and cuBLAS handles (#3081, thanks @Harshan01!)
• Use intptr_t for cuDNN handles (#3082, thanks @Harshan01!)
• Minor fix to using_allocator (#3094)
• Remove IndexOrValueError (#3096)
• Remove unused argument in fill_diagonal (#3171)
• Silence sign comparison warning (cont'd) (#3181, thanks @leofang!)
• Avoid enum comparison (-Wenum-compare) (#3182, thanks @leofang!)
• Remove unused, deprecated fields from cudaPointerAttributes (#3183, thanks @leofang!)

Documentation

• Update installation guide for conda-forge (#3052, thanks @leofang!)
• Include UnownedMemory in the API docs (#3086, thanks @jakirkham!)
• Fix gencode example in the doc (#3147, thanks @leofang!)
• Document convolve and correlate (#3161, thanks @jakirkham!)

Examples

• Add mpi4py examples (#3049, thanks @leofang!)

Tests

• Added a compute capability check for testing grid sync (#3051)
• Fix tolerance of fft tests (#3056)
• Skip irfft tests for compute capability != 7 (#3084)
• Rewrite tests not to use numpy_cupy_raises cupyx.* tests (#3099)
• Rewrite manipulation tests not to use numpy_cupy_raises (#3122)
• Remove python 2.7 builds (#3162)

Others

• Add copyright notice for Random Kit (#3107)
• Bump version to v8.0.0b1 (#3204)