Investigate recent test failures, concurrent with upgrade of DPC++ to 2025.0.1

[oleksandr-pavlyk]

Consider snippet:

import dpctl.tensor as dpt

def test(n):
    x = dpt.arange(n*n, dtype="f4")
    m = dpt.ones(n*n, dtype="?")
    return dpt.all(x[m] == x)

Then

In [38]: [n for n in range(2, 265) if not test(n)]
Out[38]: [260, 261, 262, 263, 264]

In [39]: [n for n in range(2, 265) if not test(n)]
Out[39]: [256, 258, 259, 260, 261, 262, 263, 264]

In [40]: [n for n in range(2, 265) if not test(n)]
Out[40]: [257, 258, 259, 261, 262, 263]

The result is always empty for dpctl compiled with DPC++ 2025.0.0

[oleksandr-pavlyk]

With

def test4(n):
    m = dpt.ones(n*n, dtype="?")
    cumsum = dpt.zeros(n*n, dtype="i8")
    q = cumsum.sycl_queue
    q.wait()
    num = dpt._tensor_impl.mask_positions(mask=m, cumsum=cumsum, sycl_queue=q)
    return dpt.asnumpy(cumsum - dpt.arange(1, n*n+1, dtype=cumsum.dtype))

I get

In [87]: np.nonzero(test4(257))
Out[87]: (array([513, 514, 515]),)

In [88]: np.nonzero(test4(256))
Out[88]: (array([0, 1, 2, 3]),)

In [89]: np.nonzero(test4(255))
Out[89]: (array([], dtype=int64),)

In [90]: np.nonzero(test4(258))
Out[90]: (array([1028, 1029, 1030, 1031]),)

[oleksandr-pavlyk]

The same thing with

def test5(n):
    m = dpt.ones(n*n, dtype="i8")
    cumsum = dpt.cumulative_sum(m)
    return dpt.asnumpy(cumsum) - np.arange(1, n*n+1, dtype=cumsum.dtype)

v = test5(257)
ind, = np.nonzero(v)
print(ind, v[ind])  # outputs [513 514 515] [-514 -515 -516]

[oleksandr-pavlyk]

Seem to affect all data types:

In [10]: def test5(n, dt):
    ...:     x = dpt.ones(n*n, dtype=dt)
    ...:     cs = dpt.cumulative_sum(x)
    ...:     return dpt.asnumpy(cs) - np.arange(1, n*n+1, dtype=cs.dtype)
    ...:

In [11]: np.nonzero(test5(257, np.float32))
Out[11]: (array([513, 514, 515]),)

In [12]: np.nonzero(test5(257, np.uint64))
Out[12]: (array([513, 514, 515]),)

In [13]: np.nonzero(test5(257, np.int64))
Out[13]: (array([513, 514, 515]),)

[oleksandr-pavlyk]

This failure is not related to the upgrade in the compiler, rather I introduced it recently by a code change during code maintenance: #1896