v0.18.0

Summary

This release achieves 100% compliance with Python Array API specification (revision 2024.12).

The release provides enhanced compatibility with NumPy 2.2.5. Window and mathematical routines are complemented by a set of new functions.

Moreover, it adds support to build dpnp from the source for AMD GPUs.

Details

Added

• Added implementation of dpnp.hamming #2341, #2357
• Added implementation of dpnp.hanning #2358
• Added implementation of dpnp.blackman #2363
• Added implementation of dpnp.bartlett #2366
• Added implementation of dpnp.convolve #2205
• Added implementation of dpnp.kaiser #2387
• Added implementation of dpnp.bitwise_count #2308
• Added implementation of dpnp.common_type #2391
• Added implementation of dpnp.interp #2417
• Added support to build dpnp for specified AMD GPU architecture using CodePlay oneAPI plug-in #2302

Changed

• Improved performance of dpnp.nansum, dpnp.nanprod, dpnp.nancumsum, and dpnp.nancumprod by reusing dpnp.nan_to_num function in implementation of the functions #2339
• Allowed input array of uint64 dtype in dpnp.bincount #2361
• The vector norms ord={None, 1, 2, inf} and the matrix norms ord={None, 1, 2, inf, "fro", "nuc"} now consistently return zero for empty arrays, which are arrays with at least one axis of size zero. This change affects dpnp.linalg.norm, dpnp.linalg.vector_norm, and dpnp.linalg.matrix_norm. Previously, dpnp would either raise errors or return zero depending on the parameters provided #2371
• Extended dpnp.fft.fftfreq and dpnp.fft.rfftfreq functions to support dtype keyword per Python Array API spec 2024.12 #2384
• Updated dpnp.fix to return output with the same data-type of input #2392
• Updated dpnp.einsum to add support for order=None #2411
• Updated Python Array API specification version supported to 2024.12 #2416
• Removed einsum_call keyword from dpnp.einsum_path signature #2421
• Updated dpnp.vdot to return a 0-D array when one of the inputs is a scalar #2295
• Updated dpnp.outer to return the same dtype as NumPy when multiplying an array with a scalar #2295
• Changed "max dimensions" to None in array API capabilities #2432
• Updated kernel header i0.hpp to expose cyl_bessel_i0 function depending on build target #2440
• Added MKL functions arg, copysign, i0, and inv from VM namespace to be used by implementation of the appropriate element-wise functions #2445
• Clarified details about conda install instructions in Quick start quide and README #2446
• Bumped oneMKL version up to 0.7 #2448
• The parameter axis in dpnp.take_along_axis function has now a default value of -1 #2442
• Updates the list of required python versions documented in Quick Start Guide #2449
• Updated FFT module to ensure an input array is Hermitian before calling complex-to-real FFT #2444
• Aligned black configuration with the list of supported python versions #2457
• Use pyproject.toml instead of setup.py aligning with current packaging best practices #2462
• Added a clarification to dpnp.linalg.cond docstring about its behavior with singular matrices #2460

Fixed

• Resolved an issue with an incorrect result returned due to missing dependency from the strided kernel on a copy event in dpnp.erf #2378
• Updated conda create commands build and install instructions of Quick start guide to avoid a compilation error #2395
• Added handling of empty string passed to a test env variable defining data type scope as a False value #2415
• Resolved build issues on non-Intel targets in dpnp.i0 and dpnp.kaiser #2439
• Ensure consistency in the dpnp.linalg.LinAlgError exception raised on singular input matrices for both non-batched and batched cases in dpnp.linalg.inv [#2458] (#2458)
• Updated test f/w to correct a check of array interface while converting to numpy.ndarray for comparison [#2467] (#2467)

New Contributors

• @david-cortes-intel made their first contribution in #2446

Contributors

• @AlexanderKalistratov
• @antonwolfy
• @david-cortes-intel
• @ndgrigorian
• @vlad-perevezentsev
• @vtavana

Full Changelog: 0.17.0...0.18.0