IntelPython / dpnp / 36875239920

79%

push

github

Update third party tests (#3023)

This PR refreshes the vendored third-party test suite under
`dpnp/tests/third_party/cupy` and `cupyx`: it syncs the tests with
upstream changes, migrates and expands coverage, adapts to newer
NumPy/SciPy, and makes the newly added tests portable across dpnp
devices (CPU/GPU, with and without native fp64 support). The change is
test-only — no library or runtime code is modified.

## What's included

New and expanded coverage:
- `cumulative_sum` / `cumulative_prod` (Array API) and batch-scan kernel
tests for `cumsum` / `cumprod`
- `cupy.linalg.tensordot`, `shape` argument to `reshape`, and additional
eigenvalue tests
- `cupyx.scipy.special` `erf` tests, and additional `test_search` /
`test_join` / `test_ndarray` cases

Framework and compatibility updates:
- Migrate `test_raw.py` to pytest
- Support and test NumPy 2.5 and SciPy 1.18
- Mark a globally-mutating test as thread-unsafe; reject CUB block
reduction for short (<128) partial reductions

Correctness fixes surfaced by the new tests:
- Fix `rfftn`/`fftn` shape mismatch when `s` and non-default `axes` are
given together
- Address casting a negative stride as unsigned int in dlpack; random
integer ranges should use uint internally
- Harden the custom along-axis scan against overflow; avoid an extra
host copy when uploading a non-contiguous NumPy array

Device portability for the new batch-scan tests:
- Build inputs at the device-supported float precision (float64 only
where fp64 is available, else float32) so NumPy and dpnp compute at the
same precision
- Relax `cumprod` tolerance on the float32 path (`rtol=1e-5`,
`atol=1e-6`); keep the fp64 path strict
- Mute the expected NumPy float32 overflow warning on long axes so the
infra-warnings tracker stays clean

1566 of 2924 branches covered (53.56%)

Branch coverage included in aggregate %.

26844 of 33212 relevant lines covered (80.83%)

8599.13 hits per line