Strange NumbaPerformanceWarning for numpy @ operator.

[EthanZhangYi]

I am using the Numba-0.45.1 with python-3.7.3

import numpy as np

from numba import njit


@njit
def do_dot():
    X = np.asfortranarray(np.ones((3, 4)))
    B = np.ones((4, 5))
    X[:, :] @ B[:, :]
    X @ B


if __name__ == '__main__':
    do_dot()

Run the python program above, X[:, :] @ B[:, :] will raise NumbaPerformanceWarning, while X @ B raise nothing. The behavior is very strange.

test.py:10: NumbaPerformanceWarning: '@' is faster on contiguous arrays, called on (array(float64, 2d, A), array(float64, 2d, A))
  X[:, :] @ B[:, :]
/usr/local/lib/python3.7/site-packages/numba/typing/npydecl.py:967: NumbaPerformanceWarning: '@' is faster on contiguous arrays, called on (array(float64, 2d, A), array(float64, 2d, A))
  warnings.warn(NumbaPerformanceWarning(msg))

[sklam]

Thanks for reporting this issue. This is interesting because it reveals that X[:, :], a __getitem__ on empty slices, causes the resulting array to loose its contiguous-ness .

The corresponding IR for the __getitem__:

 $0.33 = static_getitem(value=B, index=(slice(None, None, None), slice(None, None, None)), index_var=$0.32)  :: array(float64, 2d, A)

[adeak]

The problem seems to be here, where the contiguity check doesn't take into account possible trailing full slices. I was planning to fix this edge case, but then I realized that if I replace my trailing colons with an ellipsis it suddenly starts working just fine, and that's more idiomatic code anyway.

Are there any possible scenarios where this leads to a false negative for the contiguity check and one can't use an ellipsis instead? I suspect not.

[GCru]

If you do
X[2:3,3:4] @ B[2:3,3:4]
the same warning message arises.

[adeak]

If you do
X[2:3,3:4] @ B[2:3,3:4]
the same warning message arises.

@GCru slicing 3:4 in the second dimension is not contiguous (assuming C i.e. row-major memory layout), is it? In which case the warning is correct.

Edit: aaah, I missed that the slices are 1 length! OK, that's a really weird edge case :)