DeadlineExceeded failures in test_eye/test_linspace

記載された pytest コマンドを実行し、array_api_tests/test_creation_functions.py を調査して、test_eye と test_linspace に焦点を当てます。生成される次元サイズを追跡し、issue で提案されている最大サイズの制限を適用します。完了条件は、両方のテストが Hypothesis DeadlineExceeded による失敗なしで完了することです。

Running array-api-tests suite on dpctl.tensor (https://github.com/IntelPython/dpctl):

ONEAPI_DEVICE_SELECTOR=*:cpu ARRAY_API_TESTS_MODULE=dpctl.tensor python -m pytest array_api_tests/test_creation_functions.py

two test new failures consistently occur (after updating from f82c7bc8627cc2c3a44fa3e425f53a253a609aa8 to recent tip of main, i.e. 9d7777bc60d45a35124b0de084ff8edbd284c06b), e.g.:

FAILED array_api_tests/test_creation_functions.py::test_eye - hypothesis.errors.DeadlineExceeded: Test took 1470.80ms, which exceeds the deadline of 800.00ms
FAILED array_api_tests/test_creation_functions.py::test_linspace - hypothesis.errors.DeadlineExceeded: Test took 864.40ms, which exceeds the deadline of 800.00ms

Here is my understanding of what happens.

Hypothesis would generate a sufficiently large array, i.e. eye(47, 100) but testing of the output correctness is is done one element at the time (suboptimal usage for offloading libraries as it triggers 4700 submissions of small kernels) and this testing would take much longer than the creation of the array itself:

In [7]: import dpctl, dpctl.tensor as dpt

In [8]: %time eye_m = dpt.eye(47,100)
CPU times: user 3.23 ms, sys: 0 ns, total: 3.23 ms
Wall time: 2.76 ms

In [9]: %time all(eye_m[i, j] == (1 if i == j else 0) for i in range(eye_m.shape[0]) for j in range(eye_m.shape[1]))
CPU times: user 6.68 s, sys: 1.98 s, total: 8.67 s
Wall time: 2.09 s
Out[9]: True

The remedy would be to restrict the maximal dimension size.