IntelPython / dpctl / 6648408413

86%

push

github

Implements ``dpctl.tensor.logsumexp`` and ``dpctl.tensor.reduce_hypot`` (#1446)

* Implements logsumexp and reduce_hypot

* Implements dedicated kernels for temp reductions over axes 1 and 0 in contiguous arrays

* logsumexp and reduce_hypot no longer use atomics

This change was made to improve the accuracy of these functions

* Adds tests for reduce_hypot and logsumexp

* Arithmetic reductions no longer use atomics for inexact types

This change is intended to improve the numerical stability of sum and prod

* Removed support of atomic reduction for max and min

* Adds new tests for reductions

* Split reductions into multiple source files

* Remove unneccessary imports of reduction init functions

* Added functions for querying reduction atomic support per type per function

* Corrected ``min`` contig variant typo

These variants were using ``sycl::maximum`` rather than ``sycl::minimum``

* Removes _tree_reduction_over_axis

Use lambdas to ignore atomic-specific arguments to hypot and logsumexp dtype_supported functions

* Always use atomic implementation for min/max if available

For add/multiplies reductions, use tree reduction for FP types, real and complex,
to get better round-off accumulation properties.

* ``logaddexp`` implementation moved to math_utils

Reduces code repetition between logsumexp and logaddexp

---------

Co-authored-by: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>

2800 of 3300 branches covered (0.0%)

Branch coverage included in aggregate %.

19 of 22 new or added lines in 2 files covered. (86.36%)

9107 of 10586 relevant lines covered (86.03%)

8053.74 hits per line