implement dpnp.nanmedian

[vtavana]

In this PR, dpnp.nanmedian is implemented.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to 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?
• If this PR is a work in progress, are you filing the PR as a draft?

[github-actions]

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

[vtavana]

Timing for median function: dpnp shows a better performance compared to numpy, in spite of the fact that dpnp uses sorting in calculating median which is not efficient.

Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Data Center GPU Max 1100 12.60.7 [1.6.31294+9]
Intel(R) OpenCL, Intel(R) Xeon(R) Platinum 8480+ OpenCL 3.0 (Build 0) [2024.18.9.0.28_160000]

import dpnp, numpy
a = numpy.ones((16*8192, 32))
axis = 1
%timeit numpy.median(a, axis=axis)

a_cpu = dpnp.array(a, device="cpu")
%timeit dpnp.median(a_cpu, axis=axis); a_cpu.sycl_queue.wait();

a_gpu = dpnp.array(a, device="gpu")
%timeit dpnp.median(a_gpu, axis=axis); a_gpu.sycl_queue.wait();

median Function	NumPy	CPU-Xeon	GPU-PVC
(2048, 2048), axis=None	17.4 ms ± 47.5 μs	16.9 ms ± 2.12 ms	2.47 ms ± 6.8 μs
(2048, 2048), axis=0	75.1 ms ± 164 μs	20.7 ms ± 1.21 ms	2.19 ms ± 4.16 μs
(2048, 2048), axis=1	17.2 ms ± 87.6 μs	15.3 ms ± 7.45 ms	1.78 ms ± 3.49 μs
(16*8192, 32), axis=None	17.3 ms ± 49.4 μs	16.5 ms ± 357 μs	2.5 ms ± 23.6 μs
(16*8192, 32), axis=0	71.5 ms ± 190 μs	18.2 ms ± 1.89 ms	2.55 ms ± 290 μs
(16*8192, 32), axis=1	19.5 ms ± 148 μs	10.7 ms ± 2.68 ms	2.34 ms ± 5.2 μs

[vtavana]

Timing for nanmedian function: When axis is not None and there a large number of axes, dpnp is slower than NumPy. This is expected because calculating nanmedian requires looping over each axis and computing the median while ignoring NaNs in that specific axis.

Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Data Center GPU Max 1100 12.60.7 [1.6.31294+9]
Intel(R) OpenCL, Intel(R) Xeon(R) Platinum 8480+ OpenCL 3.0 (Build 0) [2024.18.9.0.28_160000]

import dpnp, numpy
a = numpy.ones((16*8192, 32))
axis = 1
# Randomly set some elements to NaN:
w = numpy.random.random((len(a.shape), 200)) * numpy.array(a.shape)[:, None]
w = w.astype(numpy.intp)
a[tuple(w)] = numpy.nan
%timeit numpy.nanmedian(a, axis=axis)

a_cpu = dpnp.array(a, device="cpu")
%timeit dpnp.nanmedian(a_cpu, axis=axis); a_cpu.sycl_queue.wait();

a_gpu = dpnp.array(a, device="gpu")
%timeit dpnp.nanmedian(a_gpu, axis=axis); a_gpu.sycl_queue.wait();

nanmedian Function	NumPy	CPU-Xeon	GPU-PVC
(2048, 2048), axis=None	19.3 ms ± 191 μs	27.1 ms ± 7.1 ms	4.38 ms ± 5.39 μs
(2048, 2048), axis=0	74.7 ms ± 66.7 μs	3.24 s ± 63.7 ms	1.23 s ± 5.14 ms
(2048, 2048), axis=1	34.1 ms ± 56.1 μs	3.21 s ± 58.1 ms	1.25 s ± 9.01 ms
(16*8192, 32), axis=None	18.7 ms ± 41.1 μs	25.5 ms ± 4.5 ms	4.41 ms ± 16.1
(16*8192, 32), axis=0	51.8 ms ± 83 μs	72.5 ms ± 4.56 ms	20.1 ms ± 215 μs
(16*8192, 32), axis=1	89.2 ms ± 265 μs	3min 31s ± 1.53 s	1min 19s ± 1.28 s

[antonwolfy]

Thank you @vtavana, LGTM!