[antonwolfy]

NumPy 2.6 adds numpy.minmax, which returns the (min, max) tuple of an array in a single pass. This PR adds dpnp.minmax with the same signature. It is backed by a new fused reduction kernel that reads the input once and writes both extrema, instead of running dpnp.min and dpnp.max one after another.

Changes

• kernels/reductions.hpp: a new tree reduction (no atomics) that keeps the running minimum and maximum per work-item and reduces both over the work-group. It uses native group algorithms for integers, all_of/any_of for bool, and NaN-aware native reductions for real floats: if any work-item holds a NaN, the NaN of the first such work-item is broadcast. Complex types use the custom SLM reduction. There are sequential, one-work-group and multi-pass variants with strided, axis-0 and axis-1 contiguous entry points, and the two outputs have independent strides.
• source/reductions/minmax.cpp: new binding _tensor_reductions_impl._minmax_over_axis(src, trailing_dims_to_reduce, dst_min, dst_max, sycl_queue, depends).
• dpnp/tensor/_reduction.py: private _minmax helper, modelled on _comparison_over_axis.
• dpnp.minmax(a, axis=None, out=None, keepdims=False, initial=None, where=True): out is a tuple of two arrays, and an entry may be None. initial and where are only supported with their default values, as in dpnp.min and dpnp.max.
• Docs, exports, changelog, and tests in test_statistics.py, test_sycl_queue.py and test_usm_type.py.

Two behaviours follow dpnp.min, dpnp.max and dpnp.modf rather than NumPy, and are not changed here:

• An out of another dtype receives the result cast after the reduction. NumPy casts the input first.
• out entries that overlap in memory are rejected, including interleaved views such as (r[:, 0], r[:, 1]).

Performance

Measured on an Intel Core Ultra 7 366H (Panther Lake H) with its integrated GPU on Level Zero and its CPU on OpenCL. The timing is the end-to-end time of dpnp.minmax(a) against dpnp.min(a), dpnp.max(a), each followed by queue.wait(). Runs are interleaved and the median is reported. The speedup is the separate time divided by the fused time; the results of both were checked to be equal in every case.

dtype	GPU 1e4	GPU 1e6	GPU 1e8	CPU 1e4	CPU 1e6	CPU 1e8
bool	1.67	1.82	2.17	2.18	3.01	4.00
int8	1.53	1.52	1.81	1.31	1.77	1.83
int32	1.24	1.23	1.55	1.59	1.11	2.03
int64	1.10	1.27	1.35	1.37	1.15	1.88
float16	1.53	1.82	2.41	2.60	3.20	4.89
float32	1.26	1.47	2.22	1.55	2.43	4.60
float64	1.28	1.36	1.41	1.47	2.74	4.69
complex64	1.42	1.47	1.53	1.21	1.12	1.10

• GPU: the fused kernel is faster in all 84 measured cases, by 1.04x to 2.88x (geomean 1.56x). For example, float32 with 1e8 elements takes 7.2 ms instead of 15.9 ms.
• CPU: the speedup ranges from 0.93x to 13.9x (geomean 2.29x). The only slower cases are int32 and int64 with 1e5 elements (0.93x and 0.97x), where dpnp.min and dpnp.max use the single-kernel atomic path, and complex128 with 1e3 elements (0.96x).
• Reductions over each axis of 2-D and 3-D arrays, and over a strided input, gain 1.2x to 2.9x on the GPU and 1.4x to 13.9x on the CPU.

Cost

Build times are for -j6 on 16 CPUs.

• The full clean build time is unchanged at 15:59. The _tensor_impl link remains the last step to finish.
• A clean rebuild of the _tensor_reductions_impl target alone goes from 488 s to 511 s (+23 s), measured branch/master/master/branch. The device link takes +20.6 s and minmax.cpp compiles in about 12 s. Compile times of the other reduction sources are unchanged.
• The _tensor_reductions_impl shared object grows from 35.38 MB to 37.27 MB (+1.89 MB, +5.4%). Of that, 1.0 MB is SPIR-V and 0.45 MB is host code. Compressed with zstd -19 it grows by 166 KB, with gzip -9 by 282 KB, and all dpnp shared objects together grow by 1.0%.
• Host RSS after import dpnp grows by about 1.6 MB. GPU memory is unchanged, and so is the first-call JIT time of the existing dpnp.min and dpnp.max.

Checklist

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• Have you added documentation for your changes, if necessary?
• Have you added your changes to the changelog?

[Copilot]

🟡 Changes recommended

The public docstring omits two intentional out incompatibilities: narrowing integer outputs can produce different extrema, and even disjoint interleaved output views are rejected.

1 open issue

• Document intentional NumPy incompatibilities for out · New

What changed in this PR

Adds dpnp.minmax to compute minimum and maximum together using a fused SYCL reduction, extending dpnp’s NumPy-compatible statistics API.

Changes:

• Adds fused kernels for contiguous and strided inputs, including multi-pass reductions.
• Integrates the public API, output handling, bindings, exports, and documentation.
• Adds tests for values, layouts, output buffers, errors, queues, and USM allocation types.

File	Description
dpnp/​tests/​test_usm_type.py	Checks output USM allocation types.
dpnp/​tests/​test_sycl_queue.py	Checks output queue placement.
dpnp/​tests/​test_statistics.py	Adds functional and edge-case tests for minmax.
dpnp/​tensor/​libtensor/​source/​reductions/​reduction_common.cpp	Registers the new reduction binding.
dpnp/​tensor/​libtensor/​source/​reductions/​minmax.hpp	Declares binding initialization.
dpnp/​tensor/​libtensor/​source/​reductions/​minmax.cpp	Adds validation, dispatch, and Python binding.
dpnp/​tensor/​libtensor/​include/​kernels/​reductions.hpp	Implements fused minimum/maximum kernels.
dpnp/​tensor/​CMakeLists.txt	Includes the new binding source in the build.
dpnp/​tensor/​_reduction.py	Adds private axis and output handling.
dpnp/​dpnp_iface_statistics.py	Adds the public API and docstring.
dpnp/​__init__.py	Exports dpnp.minmax.
doc/​reference/​math.rst	Lists minmax among extrema functions.
CHANGELOG.md	Records the new functionality.

_{Review effort: Balanced}

[github-actions]

Array API standard conformance tests for dpnp=0.22.0dev1=np2py314h8d9cdd5_24 ran successfully.
Passed: 1491
Failed: 0
Skipped: 6

[github-actions]

View rendered docs @ https://intelpython.github.io/dpnp/index.html

[coveralls]

coverage: 84.467% (+0.04%) from 84.424% — add_minmax_impl into master

[Copilot]

🟡 Changes recommended

Casting a read-only output to a different dtype hides its read-only flag, allowing the reduction to mutate the input before failing; validate original output writability before reduction and add a mixed-dtype regression test.

1 open issue

• Validate original output writability before casting · New

1 resolved since last review

• Document intentional NumPy incompatibilities for out

_{Review effort: Balanced}

[antonwolfy]

Follow-up proposal: the ASV suite has no reduction benchmarks yet, so nothing tracks the gain of the fused minmax kernel over separate min + max calls. Once this PR is merged, I plan to open a separate PR that adds benchmarks/benchmarks/bench_reductions.py, following bench_elementwise.py:

• the function as a parameter: min, max, minmax
• float dtypes and sizes from the shared axes in _utils.py
• axis=None, 0 and -1 on a 2-D input, to time both the contiguous and the strided kernels

Keeping it out of this PR means the benchmark never runs against a dpnp build without dpnp.minmax.

[Copilot]

🔵 Needs a closer look

Human assessment is needed to weigh the performance impact of unnecessary device conversions and full-output memory traffic when output dtypes differ.

No open issues 🎉

1 resolved since last review

• Validate original output writability before casting

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

No active review issues were identified for the new fused dpnp.minmax reduction.

No open issues 🎉

_{Review effort: Balanced}

[vlad-perevezentsev]

No more comments from my side
LGTM
Thank you @antonwolfy