Update reduction data types for 2023.12 array API specification, update __array_api_version__

[ndgrigorian]

This pull request updates reductions over floating point data to be performed in the input type by default as per the 2023.12 array API spec.

This pull request also changes __array_api_version__ to 2023.12. As cumulative_sum is a requirement of the 2023.12 spec, this PR should be redirected to master after the base branch is merged.

• 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?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_190 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[oleksandr-pavlyk]

I think the spec changed default output dtype of reduction operations only for the floating point input dtypes.

For integrals previous behavior is correct, but I will double-check.

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_191 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[oleksandr-pavlyk]

I think the spec changed default output dtype of reduction operations only for the floating point input dtypes.

For integrals previous behavior is correct, but I will double-check.

Yes, the change to dtype=None behavior was made in data-apis/array-api#744, The PR works as specified though:

In [13]: def check(dt): return dpt.sum(dpt.zeros((10**3, 10**2), dtype=dt, device="cpu"), axis=0).dtype

In [14]: { inp:check(ty) for inp, ty in dpt.__array_namespace_info__().dtypes().items()}
Out[14]:
{'bool': dtype('int64'),
 'float32': dtype('float32'),
 'complex64': dtype('complex64'),
 'complex128': dtype('complex128'),
 'int8': dtype('int64'),
 'int16': dtype('int64'),
 'int32': dtype('int64'),
 'int64': dtype('int64'),
 'uint8': dtype('uint64'),
 'uint16': dtype('uint64'),
 'uint32': dtype('uint64'),
 'uint64': dtype('uint64')}

[oleksandr-pavlyk]

The "array_api_tests/test_special_cases.py::test_unary[signbit(x_i is NaN) -> True]" failure seems like a bug in the test suite.

This is the output of print(case, ": ", x.shape, x.strides, x._element_offset, x.dtype, x.device, res.shape, res.dtype, res.device, x, res, idx, hex(int(x_v[idx])), in_, out, binary(in_)) added to the test_unary function:

x_i is NaN -> True :  () () 0 float32 Device(opencl:cpu:0) () bool Device(opencl:cpu:0) nan False () 0x7fc00000 nan 0.0 01111111110000000000000000000000

The "0x7fc00000" is hexadecimal view into float32, and it indeed does not have the sign bit set.

I filed data-apis/array-api-tests#250

[ndgrigorian]

The "array_api_tests/test_special_cases.py::test_unary[signbit(x_i is NaN) -> True]" failure seems like a bug in the test suite.

This is the output of print(case, ": ", x.shape, x.strides, x._element_offset, x.dtype, x.device, res.shape, res.dtype, res.device, x, res, idx, hex(int(x_v[idx])), in_, out, binary(in_)) added to the test_unary function:

x_i is NaN -> True :  () () 0 float32 Device(opencl:cpu:0) () bool Device(opencl:cpu:0) nan False () 0x7fc00000 nan 0.0 01111111110000000000000000000000

The "0x7fc00000" is hexadecimal view into float32, and it indeed does not have the sign bit set.

I filed data-apis/array-api-tests#250

Yes, there are a couple more cases too: the copysign failure also seems to be incorrect, both inputs appear to be zero but the output is expected to be NaN, and the cumulative_sum test failure comes from an assumption that cumulative_sum(M) for M size 1 and 1D should return an array with empty shape (i.e., a scalar).

The prod failure is a result of a test not yet updated for the 2023.12 spec.

If these failures are still around come time to merge this branch, I will introduce them to the array API test skips file.

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_195 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_196 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[coveralls]

coverage: 88.014% (+0.05%) from 87.967%
when pulling 0722cdf on increment-array-api-version
into 65bb9ef on master.

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_198 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[github-actions]

Array API standard conformance tests for dpctl=0.17.0dev0=py310h15de555_163 ran successfully.
Passed: 869
Failed: 9
Skipped: 92

[oleksandr-pavlyk]

For the a future PR, we could parametrize default dtype functions to take callables that return needed default type given q. In tests we can use mock functions to reach branches left out by CI runs.